1. Pick the strongest reasoning model.

   Use your assistant's strongest reasoning mode for decisions. In Claude, choose Opus and turn on Extended thinking; in another tool, use its equivalent reasoning mode. The default model is fine for drafts.

2. Attach quotes.csv, vendor_price_increase_letter.txt, and customer_emails.txt. Paste the decision.

   Prompt 1 · the blanket order decisionCopy

   I'm the GM of a hose and fittings distributor. An OEM customer, Ironvale Fabrication, accepted our blanket order proposal (quote Q-4512 in the attached quotes file) for 60 machines with monthly releases starting October, wants us to hold the 100-machine price, wants us to do the crimping, and needs a certificate of conformance with each release. His email is message 3 in the attached inbox. Our hose vendor's October price increase letter is attached too.
   
   Think this through before you answer. What's the margin risk of holding the price through a 6.5% increase? What are my options (hold and buy ahead inside the vendor's window, hold with an escalation clause, reprice for 60)? What would have to be true for each to be the right call? Then recommend one, tell me what you're unsure about, and list the three numbers I should verify by hand before I reply to Marco.
   
   Use only quotes.csv, message 3 in customer_emails.txt, and vendor_price_increase_letter.txt as evidence. Identify each source by file, record, and date when available. First restate the decision and build a facts table with quoted quantity, 60-machine release plan, 100-machine price, current component costs, stated increase, effective date, buy-ahead deadline, crimping, COC, and any cancellation language. Put “not supplied” beside missing facts; do not invent freight, carrying cost, payment terms, scrap, capacity, tax, or vendor availability.
   
   Show formulas and intermediate values for current gross margin dollars and percent, post-increase margin, inventory cash required, and exposure on unreleased material. Run low, expected, and high cases where an input is missing, labeling assumptions. Compare hold-and-buy-ahead, escalation, and reprice-for-60 in a table with economics, customer impact, operational risk, reversibility, and conditions for success. Include edge cases: delayed releases, fewer than 60 machines, cancellation, mixed pre- and post-increase receipts, certification cost, obsolete inventory, and a second vendor increase.
   
   Finish with: executive recommendation; strongest reason it may be wrong; confidence level; negotiation position and fallback; three numbers to verify; five questions for sales, purchasing, or operations; and a human approval checklist. Do not draft customer language yet. A GM must confirm costs, vendor window, COC capability, capacity, and contract terms before acting.

   Watch: the reasoning appears before the answer. Read it. That is where the assumptions live.

3. Argue the other side.

   Prompt 2 · the pushCopy

   Now argue against your own recommendation as hard as you can. Keep the Ironvale Fabrication blanket order, quote Q-4512, message 3 in customer_emails.txt, and the dated vendor increase letter as the complete evidence set. Do not introduce market facts or customer intentions that are not in those sources.
   
   Begin by restating the prior recommendation, the assumptions carrying the most weight, and the evidence supporting each. Then act as three skeptical reviewers: our CFO protecting cash and gross margin, our operations manager protecting capacity and execution, and Marco at Ironvale protecting price and supply continuity. For each reviewer, make the strongest evidence-based objection, identify the failure mode, estimate the financial or relationship consequence where the files permit it, and name the early warning signal.
   
   Build a six-month pre-mortem dated from the October release start: assume the recommendation failed and explain five plausible reasons. Include release slippage, cancellation or volume shortfall, vendor increase timing, inventory carrying exposure, crimp/COC workload, and contract ambiguity where relevant. For every reason, show which fact supports it, what remains unknown, and what mitigation could be negotiated now. Compare the best alternative against the recommendation using the same criteria and calculations; do not make the alternative look better by changing assumptions.
   
   Conclude with a “change my mind” table: evidence that would reverse the recommendation, who can provide it, and the deadline to obtain it. State whether you changed your recommendation, strengthened it, or made it conditional, and why. Give a confidence rating with a plain-language explanation. End with a human review checklist requiring finance to verify margin math, purchasing to verify buy-ahead terms and availability, operations to verify crimp and COC capacity, and sales leadership to approve customer terms before anything is promised.

4. Make it interview you.

   Prompt 3 · ask me firstCopy

   Before you write the reply to Marco at Ironvale Fabrication, use quote Q-4512 in quotes.csv, message 3 in customer_emails.txt, and vendor_price_increase_letter.txt to identify what is known and missing. Ask me exactly five decision-critical questions, one at a time, and wait after each question. Do not assume an answer or ask several questions in one sentence. Prioritize confirmed release quantities and dates, price and escalation authority, cancellation liability, lead time and buy-ahead commitment, and our ability and cost to crimp and provide a COC. If one is already answered explicitly in a source, cite it and ask the next most important unknown instead.
   
   After my fifth answer, create a short answer log separating my supplied facts from file evidence. Flag any conflict and stop for clarification rather than choosing a convenient version. Then draft two items.
   
   First, write Marco a concise, professional reply that acknowledges the October start, confirms only approved terms, makes the next action unmistakable, and does not expose our internal margin. Use specific dates and quantities. Do not claim inventory is reserved, certification is available, or a price is fixed unless the evidence supports it.
   
   Second, write a proposed blanket-order term sheet with sections for scope and kit definition; committed quantity; monthly release schedule and forecast mechanism; unit price and validity; escalation formula, trigger, notice, and evidence; crimping and COC requirements; lead time and expedite treatment; freight and taxes if known; cancellation, rescheduling, and liability for purchased material; substitutions; acceptance; and named owners. Put [CONFIRM] beside any unresolved term.
   
   Finish with a source-and-assumption table, the calculations behind any price or exposure number, edge cases the language covers, and a pre-send checklist. Require sales leadership to approve commercial terms, purchasing to confirm vendor dates and costs, operations/quality to approve crimping and COC language, and an authorized reviewer to confirm contractual wording. This is a business draft, not legal advice.

   Answer the questions the way you'd actually answer them. The reply it writes will be yours, not generic.

5. Check yourself.

   I read the reasoning, not just the answer It named what it was unsure about It asked me questions before writing the reply

I want to redesign how we [quote from emailed RFQs / review the monthly numbers / handle vendor price letters / follow up on open quotes] at a [hose and fittings distributor with 3 branches and 34 people]. Act as an operations designer who has done this for distributors and knows what AI can and can't do reliably today.

Work in four rounds and wait for me after each:

Round 1: interview me. Ask the five questions that matter most about how it works today: who touches it, what comes in, what goes out, where it breaks, what "good" looks like. One question at a time.

Round 2: map the current process as a numbered list of steps with who does each, how long it takes, and where the errors and delays are. Ask me to correct it.

Round 3: propose the future process. For each step, say whether it stays human, goes to the AI assistant with a human approving, or runs without a human. Add the guardrails: what data it can see, what it's not allowed to do, what it must flag, what a person checks before anything leaves the building. Be honest about what shouldn't be automated yet.

Round 4: write the spec: a one-page description someone could build from, the first prompt or skill for the AI step, the test we'd run for two weeks to know it works, and the three numbers we'd watch.

1. New chat. Attach the workbook. Ask what's in it.

   Prompt 1 · orientCopy

   You are a distribution data analyst orienting an owner to Sierra_Hose_Workbook.xlsx, an ERP export with one sheet per table. Do not start business analysis yet. Profile the workbook so we know whether later margin, customer, rep, inventory, AR, quote, and branch conclusions are safe.
   
   For every sheet, report exactly two concise lines: first, its business purpose, grain (what one row represents), date range, row count, and likely primary key; second, its relationships to other sheets and its most important quality warnings. Inspect headers and values rather than relying only on sheet names. Identify candidate keys and joins such as customer, SKU, quote, branch, rep, invoice, or period, but never claim a relationship is proven when key coverage does not support it.
   
   Then provide a workbook-level audit. Quantify duplicate keys, blank required fields, dates outside the apparent range, negative or zero quantities, impossible percentages, inconsistent units, inconsistent spellings/case, orphan keys, discontinued items still used, and totals that fail to reconcile. Distinguish normal business records such as credits or returns from likely defects; if the workbook does not define the treatment, say so. Include the sheet, column, affected row count, two representative record identifiers, severity, and the analysis that could be distorted.
   
   Produce: a sheet inventory; relationship map in plain text; data-quality issue table; field-definition questions for the ERP owner; and a go/caution/stop recommendation for sales and margin analysis. Cite sheet names, columns, row identifiers, and dates for every warning. Do not “clean” data silently. Finish with a human verification checklist asking the ERP owner to confirm key definitions, treatment of returns/credits/transfers, gross-margin formula, dates, and any exclusions before we trust later findings.

2. Ask the owner's question.

   Prompt 2 · find the problemsCopy

   Use the Sales_History sheet in Sierra_Hose_Workbook.xlsx. I'm the owner with ten minutes before a sales meeting.
   
   Tell me: (1) where margin is slipping and since when, (2) which customers are drifting, (3) whether any rep discounts more than the others, (4) anything else that looks off. Numbers and months, and how confident you are in each. Don't describe the data; find problems, biggest first.
   
   Before calculating, state the sheet’s date range, row grain, revenue, cost, margin, discount, customer, rep, branch, and category fields. Explain how you handle returns, credits, zero-revenue rows, blanks, and duplicate transaction IDs. Do not silently exclude records. Reconcile analyzed revenue and gross margin dollars to sheet totals.
   
   Define “slipping” and “drifting” before using those words. Compare recent periods with a sensible prior period and show both dollars and percentages so a small account does not outrank a material one. For margin, calculate gross margin dollars and gross margin percent from source fields and show formula, baseline, current result, change, first month of sustained decline, and revenue affected. For customer drift, report last purchase date, recent versus baseline revenue, categories affected, and whether seasonality or a one-time order could explain the signal. For rep discounting, compare weighted averages and mix-adjusted results if the data permits; do not accuse a rep based on simple averages.
   
   Rank findings by estimated business impact. Each finding must include the exact number, months, customer/rep/category/branch, supporting row or filter definition, confidence, alternate explanation, and question that would confirm it. Include at least one challenge test that could disprove each of the top findings. Put data defects in a separate section rather than presenting them as performance issues.
   
   Output an owner brief with: five-line executive summary; ranked findings table; supporting calculation table; three meeting questions; three actions with owners; uncertainties; and a human verification checklist. Require finance or the ERP owner to validate totals and definitions, and the sales manager to review account context before any personnel or pricing decision.

   Check: four problems are planted. How many did it find? Then check one number yourself: filter industrial hose, February against March.

3. Build the heat map and the chart.

   Prompt 3 · make it visualCopy

   Using only the verified Sales_History sheet in Sierra_Hose_Workbook.xlsx, create two decision-ready visuals for the owner’s sales meeting. Preserve the workbook’s actual date range and do not fabricate missing months or categories.
   
   First build a customer-by-product-category heat map. Customers go down the side, categories across the top, cell color represents weighted gross margin percent, and the printed value in each cell is revenue. Calculate gross margin percent as total gross margin dollars divided by total revenue for the cell, not an average of row percentages. Sort customers by total revenue descending. Include a total row and column, clear currency and percent formatting, a colorblind-safe scale with labeled endpoints, and a visible treatment for no sales, zero revenue, returns, and insufficient data. If there are too many customers for a readable slide, show the top group plus “all other,” state the cutoff, and provide the complete table separately.
   
   Then build a monthly line chart of weighted gross margin percent by category. Keep one consistent color per category, label the date range, show the company total as a reference, and annotate only material turning points supported by the data. Do not imply causation. If sparse volume makes a point unstable, mark it and disclose the revenue denominator.
   
   Deliver each visual with a descriptive title that states the finding, a one-sentence reading guide, the single insight visible here that was hard to see in the table, and one question for the room. Also provide the aggregation table behind each chart, formulas, filters, excluded records, and a reconciliation to Sales_History totals. Check labels, scales, sort order, and arithmetic. End with a human review step: the analyst verifies two cells and two monthly points against source rows, confirms that returns/credits are handled correctly, and approves the visuals before they enter a slide.

   The heat map appears in a panel beside the chat. Click it. Ask for changes: "Only customers over $20K." "Add a column for the trend."

4. Turn it into the deck.

   Prompt 4 · the PowerPointCopy

   Turn the verified analysis from Sierra_Hose_Workbook.xlsx into an editable six-slide PowerPoint for Monday’s sales meeting. The audience is branch and sales leadership with limited time. Use only findings already supported by Sales_History; do not add explanations, targets, or forecasts that were not established.
   
   Slide 1 is a plain title slide with the actual analysis period and “Monday Sales Meeting.” Slides 2–5 each cover one ranked finding. Every finding slide must have a conclusion-style title, one large decision-relevant number, the chart or compact table that proves it, the comparison period and denominator, a one-sentence evidence note naming the sheet and filter, and one question for the room. Slide 6 lists exactly three decisions, each with proposed owner, needed evidence, and decision deadline.
   
   Prefer charts already created in the prior step; do not redraw them with changed scales or calculations. Keep numbers large, labels readable, colors restrained and consistent, and whitespace generous. No clip art, decorative stock photos, unsupported icons, or paragraphs. Use currency and percentages consistently. Add short speaker notes to slides 2–6 explaining what is fact, what is interpretation, what remains uncertain, and the source rows or aggregation needed to reproduce the number.
   
   Before generating the .pptx, show me a slide plan and a source ledger mapping every displayed number to workbook sheet, fields, filters, and formula. Flag any unresolved quality issue rather than hiding it. Create the file only after the plan is internally consistent. Then run a quality check: totals reconcile, dates match, no text is clipped, charts have labeled axes, claims do not exceed evidence, and all six slides are editable. Give me the file and a final human review checklist for the owner and data analyst before Monday.

   The AI assistant produces a real .pptx you can download and open. Then: "Also give me an Excel file with a Findings sheet holding each table you used."

5. Save it as the monthly pattern.

   Prompt 5 · make it repeatableCopy

   Write reusable operating instructions for a monthly Sierra Hose sales review. On the first of every month, I will attach a new ERP workbook in the same format as Sierra_Hose_Workbook.xlsx. The process must orient to the workbook, run the same four checks, compare results with the prior month, build the customer-category heat map and category margin trend, draft the six-slide deck, and give me a hand-check list.
   
   Write this as a briefing a capable analyst can follow without seeing today’s chat. Include: purpose and audience; required file and expected sheets/columns; reporting cutoff and comparison periods; definitions and formulas for revenue, gross margin dollars, weighted margin percent, customer drift, sustained margin slip, and rep discounting; handling for duplicates, blanks, returns, credits, transfers, zero revenue, new customers, closed accounts, missing months, and schema changes; and explicit stop conditions.
   
   Specify the sequence as numbered stages: validate the file; reconcile totals; run four analyses; challenge each finding; create visuals; assemble the deck; and request approval. Define the exact output structures, file naming convention with reporting month, source ledger, and issue log. Require citations to sheet, field, record/filter, and month for every material claim. Never silently clean data, infer causes, average percentages incorrectly, or compare incomplete months with full months.
   
   Add acceptance tests with known checks from the current workbook, including two manually recalculated heat-map cells and two monthly category points. Explain how to compare with the last approved run without carrying forward stale conclusions. End with roles: ERP owner verifies definitions and extract completeness; finance verifies reconciliations and formulas; sales leadership validates customer and rep context; presenter approves the final deck. If a check fails, produce an exception report and no presentation-ready conclusion.

   New Project called "Monthly numbers." Paste that as the instructions. Next month: attach the file, say "run it." Capability 8 shows how.

6. Check yourself.

   It named the industrial hose margin drop and the month A heat map I changed by asking A PowerPoint file on my laptop Intermediate: a "Monthly numbers" Project exists

1. The deck. Attach branch_pnl.csv.

   Prompt 1 · branch review deckCopy

   You are preparing a quarterly operating review for the owner of Sierra Hose. Use only branch_pnl.csv, which contains monthly results for three branches over the stated year. First profile the file and confirm the exact date range, branches, units, row count, and definitions available. Recalculate revenue, gross margin percent, EBITDA, EBITDA percent, freight-out percent, and per-head measures where the required fields exist. Show formulas, reconcile annual branch totals to the source, and flag missing months, duplicate branch-months, signs, or percentages that do not match their dollar fields.
   
   Build an editable PowerPoint: title and period; one slide per branch; one comparison slide; and one questions slide. Each branch slide must show revenue, weighted gross margin percent, and weighted EBITDA percent for the review period, changes from the prior comparable period available in the file, one thing working, one thing not working, and one specific priority for the quarter. Support every interpretation with a compact chart or table and a source note naming branch, months, fields, and formula. Do not invent budget, target, cause, or manager performance. Treat unusual freight, labor, occupancy, headcount, or mix as a hypothesis and phrase the question needed to confirm it.
   
   The comparison slide should use consistent scales and definitions across all three branches and distinguish dollar size from rate performance. The final slide should organize questions by branch and include three cross-branch decisions. Use conclusion-style titles, large numbers, plain language, restrained colors, and no clip art. Add speaker notes separating fact, interpretation, uncertainty, and follow-up.
   
   Before creating the file, provide a slide plan and source ledger. Then generate the .pptx and check that charts are editable, labels are readable, dates and totals agree, no claim exceeds the evidence, and no text is clipped. Give me the file plus a human review checklist for finance and each branch manager before the deck is presented.

   Then edit: "Boise's freight line doubled in the spring. Make that the headline of the Boise slide and add a chart of freight_out % by month."

2. The document. Attach vendor_price_increase_letter.txt.

   Prompt 2 · the customer notice as a Word fileCopy

   Act as the commercial communications lead for Sierra Hose. Read vendor_price_increase_letter.txt as the authoritative source and create an editable Word notice for our top 25 accounts. Extract the vendor’s letter date, effective date, affected categories, stated percentages, tariff treatment, order deadline, and any exclusions or conditions. Build a source table first. If the attached letter does not support one of the requested claims, mark it [CONFIRM] and do not present it as fact.
   
   The customer message should state that our adjustment is effective October 1, hydraulic hose will increase about 6%, industrial hose about 4%, Sierra Hose will absorb the tariff surcharge through year end, and qualifying orders placed by September 20 ship at current pricing, but only to the extent the source and our supplied policy support those statements. Make the September 20 fall-order request the single clear action. Include one credible line about why customers buy from us without inventing service guarantees, inventory commitments, or savings.
   
   Deliver: an email subject line; a brief email body; and a one-page letter with date, greeting suitable for mail merge, direct explanation, category table if useful, action and deadline, contact path, and professional close. Tone is calm, factual, and confident: no apology spiral, blame, pressure tactics, or unsupported macroeconomic claims. Preserve customer confidentiality and do not name the vendor unless instructed.
   
   Create a clean, editable .docx using a restrained letterhead style with a placeholder for the approved logo and contact details rather than fabricating them. Check dates, percentages, page breaks, accessibility, and consistency between email and letter. Finish with an approval checklist requiring purchasing to verify vendor terms, finance to approve customer adjustments and surcharge treatment, sales to confirm the account list, and an authorized manager to approve wording before release.

3. The one-pager. Suppliers especially.

   Prompt 3 · counter launch kitCopy

   We are launching [product line, or use "a new static-dissipating petroleum transfer hose"] to distributor counters. Act as a supplier product trainer. Use only the attached approved specification sheet and any product information I provide in this chat. Identify document title, revision, and date. Do not infer chemical compatibility, pressure, temperature, conductivity, approvals, coupling compatibility, or applications beyond that evidence.
   
   Create one printable, editable document for a counter person who has 30 seconds with a customer. Page one is a cheat sheet: what the product is in one sentence; ideal customer and use cases; three discovery questions; verified operating limits; two products or practices it may replace and the conditions under which that comparison is valid; one common objection and a factual answer; and a prominent “stop and verify” box. Include a short selection path, but route any unsupported size, media, pressure, temperature, static-control, or fitting question to the approved chart or product specialist.
   
   Page two contains a five-question practical quiz with realistic counter situations, an answer key, and one-sentence rationale citing the specification section. Include at least one case where the correct answer is not to recommend the product. Finish with a short email a branch manager can forward that explains the customer problem, what is new, a two-minute action for the counter team, and where to find approved technical help.
   
   Keep language plain, layouts scannable, type large enough to print, and claims faithful to the source. Separate benefits from specifications. Add revision date, document owner, and “verify current specifications before quoting” footer. Before finalizing, provide a claims ledger and uncertainty list. Require product engineering or product management to verify technical claims and sales leadership to approve comparisons and launch language.

4. Teach it your voice, then redo one.

   Prompt 4 · my voiceCopy

   Here are five emails I have sent with names and sensitive details removed: [paste]. Act as an editorial analyst, not an impersonator. Identify repeatable characteristics of my business writing from this evidence: opening style, sentence length, directness, vocabulary, use of bullets, treatment of bad news, requests, closings, and phrases I avoid. Quote brief examples from the five samples and note when the evidence is mixed. Do not infer personality, age, background, or preferences that the samples do not establish.
   
   Produce a voice profile with: five short rules suitable for pasting into future prompts; three “do” examples; three “do not” examples; and a confidence note for each rule. Distinguish my voice from accidental features of a single email. Preserve professional clarity even if a sample is vague or contains a mistake. The rules must never override factual accuracy, legal review, customer respect, or the approved content of a source document.
   
   Then rewrite the October customer price-adjustment notice created from vendor_price_increase_letter.txt. Preserve every approved date, category percentage, tariff statement, deadline, qualification, and customer action exactly; change only style and organization. Provide a change log showing how each of the five voice rules affected the draft. Do not introduce promises, soften conditions, name confidential parties, or make unsupported claims to sound more like me.
   
   Return the revised email subject, email body, and one-page notice, followed by a fact-preservation checklist comparing old and new versions. Flag any conflict between “my voice” and clear customer communication. End with human review: I compare the draft with the five source emails, purchasing and finance recheck commercial facts, and the authorized sender approves the final language.

   Save the five rules as a skill called "My voice." Every document after this sounds like you.

5. Check yourself.

   A .pptx on my laptop that I changed by asking A .docx notice with one clear ask Intermediate: five voice rules saved

1. New chat. Attach rfq_emails.txt and price_list.csv. Paste the prompt.

   Prompt 1 · build the quoteCopy

   I run quoting at a hose and fittings distributor. Attached is our price list and three RFQ emails. Start with RFQ 1 from Red Rock Aggregates.
   
   Build the quote: match each line to a SKU, work out quantities (hose per foot, fittings each, hydraulic assemblies need two fittings, two ferrules, and crimp labor per end), and give me a clean quote table at list.
   
   Where the customer is vague, do not guess silently. List the questions I should ask Dale first. Then suggest two or three add-ons a wash plant would need.
   
   Treat rfq_emails.txt and price_list.csv as the only evidence. Cite RFQ line/message and price-list SKU for every match. Start with an extraction table containing customer wording, requested quantity, unit, application, size, material, ends, pressure, temperature, and need date. Preserve Dale’s wording. Mark missing or ambiguous requirements and separate required clarification from optional commercial questions.
   
   Build a candidate-match table with requested line, proposed SKU, exact description, quantity math, unit, list price, extended price, match confidence, and evidence. Hydraulic assemblies require hose length converted to feet using the stated selling unit, two compatible fittings, two ferrules, and SVC-CRIMP for each end. Show every formula and do not substitute a size, end, material, or series because it seems likely. If no exact supported match exists, say “no supported match” and stop that line.
   
   Draft the quote table only for resolved lines. Put unresolved lines in a hold section, not a priced guess. Include validity, availability, lead time, freight, tax, certification, and terms only when the files supply them; otherwise label [CONFIRM]. Suggest two or three wash-plant add-ons as questions, not assumed requirements, and tie each to evidence in the catalog.
   
   Finish with questions for Dale ranked by what blocks the quote, a source ledger, edge-case check for pack quantities and minimum orders, and a human review checklist. An experienced inside seller must verify SKU compatibility, quantities, price-list date, stock/lead time, and application suitability before sending.

   Check: did it ask about "the aluminum ones we always get," "-08 I think," and which cams go on the 2" hose? If it guessed: "You guessed on the coupling type. Ask me instead."

2. Push it: the OEM kit and a volume tier.

   Prompt 2 · per-kit price and a tierCopy

   Now use RFQ 3 from Ironvale in rfq_emails.txt and the dated price_list.csv. Act as the quote-desk lead. Extract the complete bill of material for one machine before pricing. Preserve requested series, sizes, lengths, ends, quantities, crimping, testing, certification, delivery, and any blanket-release language. Cite the RFQ text and SKU row supporting every selection; unresolved compatibility or configuration stays [CONFIRM].
   
   Price one kit and 25 machines at list. Show line-by-line SKU, description, per-kit quantity, purchasing/selling unit, unit cost, list price, per-kit extension, 25-machine extension, and formula. For every assembly include hose footage under the stated rounding rule, two correct fittings, two correct ferrules, and two SVC-CRIMP labor charges unless the source explicitly says otherwise. Explain Marco’s stated reason for asking us to crimp; if the RFQ gives no reason, label your explanation a hypothesis and ask him.
   
   Then design a 100-machine tier that protects at least 30% gross margin on the complete kit. Use sell price = cost / (1 - margin), not markup. Show total cost, minimum sell, proposed rounded sell, margin dollars and percent at one, 25, 60, and 100 machines. Do not invent volume cost breaks. Run sensitivity for the vendor increase if it affects these components and identify what date or purchasing assumption drives the result.
   
   Include edge cases: partial releases, mixed cost receipts, scrap/rework, testing or COC, freight, cancellation, and component minimums. Output a customer-facing table, internal margin table, assumptions/open questions, and negotiation guardrails. Purchasing verifies current cost and availability, operations verifies labor and capacity, quality verifies testing/COC, and a manager approves the tier before it is quoted.

3. A photo. Anything on your phone.

   Take a picture of a hose tag, a nameplate, a handwritten pick list, a page of a catalog. Attach it. "What is this, what would I quote for it, and what should I ask before I do?" Photos of the real world are files too.

4. Save the pattern.

   Prompt 3 · make it repeatableCopy

   Turn the Red Rock and Ironvale quote work into reusable instructions for the top of a new chat. Write a briefing for a capable new inside sales hire who will attach a current price list and one or more RFQs. The instructions must work without access to today’s conversation and must never assume the same customer, product, or date.
   
   Define the job, required files, accepted columns, source priority, and stop conditions. Give a numbered workflow: identify file dates and versions; extract customer requirements verbatim; normalize quantities and units without losing original wording; identify ambiguities; match only supported SKUs; calculate assemblies and kits; price at list; calculate any tier with gross-margin formulas; suggest evidence-based add-ons; and prepare customer and internal outputs.
   
   Specify exact tables for requirements, questions, SKU matches, quote lines, margin analysis, and sources. Require two fittings, two compatible ferrules, and crimp labor per end only when the assembly and catalog support them. Cover pack quantities, minimums, superseded items, unavailable combinations, unit conversion, rounding, mixed costs, volume releases, testing/COC, freight, lead time, and cancellation. Never silently guess a SKU, compatibility, stock, cost, lead time, or customer intent. Distinguish fact, assumption, and recommendation.
   
   Add acceptance tests based on known Red Rock and Ironvale examples without hard-coding their answers, plus failure tests for vague descriptions and missing catalog combinations. Include file naming and revision-date rules. Finish with the approval chain: inside sales checks requirements and arithmetic; product/operations checks application and assembly; purchasing verifies cost and availability; a manager approves below-floor or tier pricing. If any required check fails, output a hold list rather than a send-ready quote.

   Project "Quote desk": those instructions plus price_list.csv attached. New chat inside it, attach only an RFQ, say "quote RFQ 2." Invite your inside sales team.

5. Check yourself.

   A quote table for Red Rock with a questions list It asked instead of guessing at least once Per-kit price and a 100-machine tier with the math visible Intermediate: a "Quote desk" Project exists

1. New chat, web search on. Pick a real company.

   Suppliers: a distributor you'd like to grow. Distributors: a plant, a contractor, an OEM. Search is on by default on paid plans; on free, add "search the web" to the prompt.

   Prompt 1 · call briefCopy

   I'm a [sales exec at a hose and fittings supplier / owner of a hose distributor] with a call next week at [company, and city if ambiguous].
   
   Research them on the web: what they do, locations, recent news, expansions, leadership changes, industries served, anything that hints at what they'd buy from me.
   
   One-page call brief: who they are in three sentences, three things happening now, three questions that show I did my homework, the one thing to offer. Link every fact. If you can't verify something, say so.
   
   Set the research cutoff to today and identify the exact company before proceeding; if the name is ambiguous, ask me to confirm the city, website, or legal name. Prioritize the company’s own site, filings, government/project records, and reputable dated trade or local news. Do not treat search snippets, lead databases, job posts, or social posts as proof without opening and evaluating the source. For every claim provide title, publisher, publication date, direct link, and access date.
   
   Separate verified facts, reasonable hypotheses, and unknowns. Research what the company makes or operates, locations, leadership, served industries, recent projects or changes, and signals relevant to industrial hose, fittings, gaskets, fluid handling, maintenance, shutdowns, safety, or distribution. Do not infer a purchase need merely from an industry label. For each potential need, state the evidence, confidence, and a question that would validate it.
   
   Structure the brief as: identity check; three-sentence company overview; dated timeline; three current signals; likely operational implications; three informed discovery questions; one evidence-based offer and one low-risk fallback; people/roles to confirm; and “do not assume” box. Keep it to one page plus sources. Include negative findings and conflicts between sources.
   
   End with a pre-call verification step: the account owner checks CRM/ERP history, confirms that contacts still hold their roles, opens the most important links, and approves the offer. Make clear that public research can inform a conversation but cannot establish private budgets, projects, incumbent suppliers, or buying intent.

   Check: click two citations. If one is off, tell it.

2. Inside plus outside. Attach account_snapshot.txt.

   Prompt 2 · the account planCopy

   Act as an account-planning partner for a hose and gasket distributor. Use account_snapshot.txt as the authoritative internal record for this dairy account and its stated late-October shutdown. State the snapshot date and data period. Separate account facts from general dairy-processing knowledge and from hypotheses that must be tested with the customer.
   
   First create a fact base: locations and contacts, purchase history by category and date, service issues, open quotes, last-year rejection of spiral wound gaskets, stated reason, shutdown timing, and any gaps. Cite the exact snapshot section for every internal claim. Do not invent plant equipment, media, temperatures, sanitation standard, gasket material, incumbent supplier, budget, or shutdown scope.
   
   Develop a 90-day plan backward from the supplied shutdown date. For wash-down hose and gaskets, list plausible use cases, failure or compliance concerns, evidence needed, discovery questions, and who should validate selection. Treat product compatibility and sanitary requirements as technical decisions requiring approved specifications. For the gasket objection, calculate no savings unless the snapshot supplies costs. Offer a value conversation that considers service life, downtime, installation, availability, and total cost while respecting that “too expensive” may still be the correct customer judgment.
   
   Output: one-paragraph objective; account fact table; opportunity/hypothesis table with confidence; week-by-week action plan; stakeholder map; opening line; eight sequenced questions; response to the prior price objection; one primary and one fallback next step; risks; and success measures. Include a “what would disprove this plan” section.
   
   End with human verification: the rep confirms dates and history in the ERP, product/application specialists approve technical statements, purchasing confirms availability and lead times, and the customer validates shutdown scope before any recommendation or commitment.

3. Deep research. Click Research, paste, and let it run.

   The Research button (next to the paperclip) runs a long, multi-source report. Start it, then do capability 6 while it works.

   Prompt 3 · market read (Research mode)Copy

   Research the industrial hose, fittings, and gasket market in [your region] for the next 12 months. Act as an independent market analyst serving both distributors and manufacturers. Define the geographic boundary and research cutoff date before searching. If “region” is ambiguous, ask for states or metros. Distinguish demand for these products from broad industry growth.
   
   Evaluate end-user segments relevant to the region, including only those supported by evidence. Look for dated projects, plant expansions or closures, maintenance cycles, public capital plans, permits, regulation, tariffs, commodity conditions, and employment or production indicators. Prefer primary sources, government data, company announcements, filings, and reputable trade reporting. Open sources rather than citing snippets. For every factual claim provide publisher, title, publication date, direct link, and access date; identify paywalls or inaccessible evidence.
   
   Create a segment table with current regional footprint, 12-month direction, strength of signal, demand mechanism for hose/fittings/gaskets, timing, evidence, counter-evidence, and uncertainty. Label a signal strong only when multiple independent sources or a direct project record support it. Do not turn national forecasts into local facts, announcements into completed projects, or one data point into a trend. Include at least one downside case and identify what would invalidate each top signal.
   
   Then recommend where a distributor might investigate adding lines, inventory, technical capability, or reps, and what manufacturers should expect distributors to request. Treat each as a testable option, not a directive. Include required investment, leading indicator, first customer interviews, and risk.
   
   Deliver an executive summary, methodology, evidence table, recommendations, watch list, and source appendix. End with a human review: local sales leaders challenge account-level assumptions, finance tests economics, product leaders verify portfolio fit, and decision owners confirm major sources before allocating people or inventory.

4. How AI describes you.

   Prompt 4 · the buyer's viewCopy

   Search the web as if a plant maintenance manager in [state] asked where to buy industrial hose, fittings, and gaskets with same-day assembly. First answer neutrally, using the same criteria for every provider: verified geography, product coverage, assembly capability, stated hours, emergency service, technical evidence, and current contact information. Do not favor [my company]. Cite every provider claim with a dated direct link and say when same-day service cannot be verified.
   
   Then audit whether [my company] appears and whether the site at [url] gives a buyer or AI system enough evidence to recommend it. Confirm the exact business and locations. Review public pages for product specificity, assembly services, service area, hours, inventory claims, technical content, certifications, location/contact consistency, structured organization information, crawlable text, and recent proof. Do not assume rankings are stable or that your search represents every buyer.
   
   Deliver: neutral buyer answer; comparison table; our visibility result; evidence that helps; evidence missing or contradictory; five prioritized site changes; and a 30/60/90-day validation plan. For each change state the buyer question it answers, exact page or content needed, evidence the business must supply, owner, effort, and measure. Do not recommend fabricated reviews, doorway pages, keyword stuffing, unsupported service claims, or content written only for machines.
   
   Include uncertainty about personalization, location, search engine, and date. End with human verification: operations confirms actual same-day capability and cutoffs, branch managers confirm addresses/hours/phones, product leaders validate technical claims, and marketing measures impressions, qualified inquiries, and conversions before declaring improvement.

   That is what your buyers see when they ask AI first. It is also your marketing plan.

5. Check yourself.

   A one-page brief on a real company with links I clicked A call plan that handles the gasket objection Intermediate: a Research report running or done

1. Attach call_transcript_sawtooth.txt. Pull the commitments.

   Prompt 1 · what did we promiseCopy

   Act as a sales-operations reviewer. Analyze call_transcript_sawtooth.txt, a call between rep Dana and a customer lost in February whom we are trying to win back. Use only the transcript. State the call date, participants, transcript quality, and any speaker-attribution uncertainty before drawing conclusions.
   
   Produce five sections. First, a commitment checklist with exact commitment, speaker, beneficiary, due date, condition, transcript timestamp or line reference, and status. Separate firm promises, intentions, requests, and ideas; never convert “I’ll try” into a commitment. Second, a table of every number: price, quantity, date, lead time, invoice amount, disputed amount, or percentage, preserving units and context. Flag conflicting or incomplete numbers and do not resolve them from memory.
   
   Third, summarize what the customer cares about using short direct quotations with references, then distinguish stated priorities from your interpretation. Fourth, assess each risk in Dana’s promises: authority, feasibility, date, dependency, margin, service recovery, and ambiguity. State evidence and the internal owner who must validate it. Fifth, identify the one issue bigger than this call, but label it a hypothesis and show what evidence would confirm or disprove it.
   
   Add an action register with owner, due date, dependency, source quote, and approval needed. Include a “not found in transcript” list for facts a manager might expect but the call never established. Do not praise or criticize Dana yet.
   
   End with a verification procedure: Dana listens to ambiguous audio or reviews the transcript, finance checks the disputed invoice and all prices in the ERP, operations confirms dates and capacity, and the account owner approves the commitment record before it enters CRM or reaches the customer.

2. Score the call, then write the follow-up.

   Prompt 2 · coaching and the emailCopy

   Using call_transcript_sawtooth.txt and the verified commitment/number tables from the prior step, act as Dana’s sales manager. Score discovery, handling the price objection, owning the February failure, asking for the whole program, and establishing next steps. Define a 1–5 rubric before scoring. For each area provide score, one sentence of evidence with timestamp or line reference, one thing done well, one missed opportunity, and a specific alternative phrase. Do not infer tone that the transcript cannot carry or punish Dana for facts she could not know.
   
   Identify the single behavior that would most improve Dana’s next call and explain why it has more leverage than the others. Add a short practice drill and success criterion. Challenge your own coaching by naming one reasonable alternate interpretation of the customer’s response.
   
   Then draft the follow-up email Dana should send Ray tonight. Keep it short and warm. Include only commitments actually made, each with owner and specific date; acknowledge the February failure without relitigating it; preserve the verified prices and disputed invoice amount exactly; state the immediate next step; and ask the one unresolved question Dana needs answered from Travis. Do not promise availability, credit, pricing, or recovery beyond the approved record. Put [VERIFY] beside anything still uncertain rather than smoothing it over.
   
   Return: rubric and scorecard; coaching priority; practice drill; email subject and body; and a fact-to-source table for every date, number, and promise in the email. Finish with human review: Dana compares the draft to the transcript, finance checks invoice and price facts, operations validates dates, and Dana or her manager approves before sending.

3. The voicemails. Attach voicemails.txt.

   Prompt 3 · three voicemails to three actionsCopy

   Act as the Reno branch inside-sales lead. Review all three records in voicemails.txt from Monday, August 24, 2026. Treat automatic transcription as imperfect. Do not invent a caller name, company, phone digit, part number, size, quantity, pressure, application, or deadline when audio text is unclear.
   
   For each voicemail create a triage row with record number, caller and company as heard, callback information, exact request, quoted phrases, date/time, urgency, urgency reason, proposed owner/role, one-line next action, missing information, and confidence. Separate customer-stated urgency from your operational judgment. Flag safety-critical application details, breakdowns, promised dates, credit issues, or product-selection questions for appropriate review. If two messages might refer to the same matter, show the evidence rather than merging them.
   
   Rank callback order and explain the tradeoff. Draft a concise callback opener and the first three questions for each caller; do not draft a product recommendation unsupported by a catalog or application evidence. Identify the new-customer opportunity a busy counter might miss, citing the words that signal it. Propose what to do today, including owner, deadline, information to capture, and a low-pressure next step. Do not manufacture account value or buying intent.
   
   Output an exception list for unintelligible or ambiguous transcript portions and a CRM-ready note that clearly labels unverified details. Finish with human verification: a counter employee checks caller details against the phone system or audio, confirms product/application facts with the caller, checks ERP/account/stock information where needed, and approves any promise before calling back.

4. Dictate. On your phone, right now.

   Open your AI assistant's app, tap the microphone, and talk for sixty seconds about a customer problem you're dealing with this week, the way you'd tell a colleague in the truck. Then: "Turn what I just said into a short email to the customer, and a separate note to my inside rep with what I need from them." Rambling in, clean out.

5. Make it a habit for every call.

   Prompt 4 · the call debrief skillCopy

   Turn the Sawtooth call analysis into reusable instructions for any future sales call or meeting transcript. Write a briefing that works without today’s chat and stays under one printed page. The user may paste a transcript or attach a file; require call date, participant names/roles if known, and approved examples of the sender’s voice if an email is requested.
   
   Define a fixed sequence: assess transcript quality and speaker attribution; extract commitments; extract every number; quote what each party cares about; identify promise and execution risks; identify the larger account issue as a labeled hypothesis; create actions; draft follow-up. Specify exact columns for the commitment checklist and numbers table, including timestamp/line citation, speaker, owner, date, condition, unit, context, confidence, and verification status.
   
   The instructions must distinguish commitments from intentions, requests, and ideas; facts from interpretations; customer statements from employee statements; and transcript uncertainty from business uncertainty. Never fill inaudible text, correct a disputed number, infer authority, or create a due date. Cover edge cases: missing timestamps, mislabeled speakers, relative dates, conflicting numbers, side conversations, multiple customers, and no clear next step. Require [VERIFY] markers and questions when evidence is incomplete.
   
   For the follow-up email, preserve approved commitments and dates, include unresolved questions, and use only supplied voice rules. Require all disputed numbers, prices, invoices, quantities, and dates to be checked against the ERP or other system of record before sending. End with an approval workflow: meeting owner checks transcript fidelity, functional owners validate promises, finance checks money, and the sender approves the email. If verification fails, produce a draft and exception list, never a send-ready message.

   Save it as a skill called "Call debrief." With your meeting tool connected (capability 9), it can run after every call without pasting.

6. Check yourself.

   A commitments checklist with dates (Thursday 9 AM, quote Wednesday, the credit) The follow-up email I'd send after a light edit I dictated something and got a clean email back

1. New chat. Paste the whole prompt. The price list is inside it.

   Build · hose assembly quote calculatorCopy

   Build me a simple web app for quoting hydraulic hose assemblies at the counter. One page, works on a phone, clean, nothing fancy.
   
   Inputs: hose series (100R1AT, 100R2AT, 100R12, 100R17), size (-04 to -16, only sizes that exist for the series), overall length in inches, end A and end B (JIC straight, JIC 45, JIC 90, Male NPT, ORFS; only what exists for the size), quantity, target gross margin % (default 35).
   
   Per assembly: hose cost = feet rounded up to the next half foot times cost per foot; plus two fittings, two ferrules, crimp labor for two ends; optional test certificate. Show cost, sell at target margin, extended total, margin dollars.
   
   "Add to quote" builds a multi-line quote with a running total. "Copy quote" produces clean plain text for an email, with lines like "SAE 100R2AT -08 x 72 in, JIC swivel straight x JIC 90, crimped and tested".
   
   Costs (SKU, description, unit, unit cost):
   HH-R2-04, SAE 100R2AT 2-wire hydraulic hose 1/4", FT, 1.7
   HH-R1-04, SAE 100R1AT 1-wire hydraulic hose 1/4", FT, 1.32
   HH-R17-04, SAE 100R17 compact 1-wire hydraulic hose 1/4" (3000 psi all sizes), FT, 1.56
   HH-R2-06, SAE 100R2AT 2-wire hydraulic hose 3/8", FT, 2.15
   HH-R1-06, SAE 100R1AT 1-wire hydraulic hose 3/8", FT, 1.68
   HH-R17-06, SAE 100R17 compact 1-wire hydraulic hose 3/8" (3000 psi all sizes), FT, 1.98
   HH-R2-08, SAE 100R2AT 2-wire hydraulic hose 1/2", FT, 2.84
   HH-R1-08, SAE 100R1AT 1-wire hydraulic hose 1/2", FT, 2.22
   HH-R17-08, SAE 100R17 compact 1-wire hydraulic hose 1/2" (3000 psi all sizes), FT, 2.6
   HH-R2-12, SAE 100R2AT 2-wire hydraulic hose 3/4", FT, 4.36
   HH-R1-12, SAE 100R1AT 1-wire hydraulic hose 3/4", FT, 3.4
   HH-R17-12, SAE 100R17 compact 1-wire hydraulic hose 3/4" (3000 psi all sizes), FT, 4.01
   HH-R2-16, SAE 100R2AT 2-wire hydraulic hose 1", FT, 6.72
   HH-R1-16, SAE 100R1AT 1-wire hydraulic hose 1", FT, 5.24
   HH-R17-16, SAE 100R17 compact 1-wire hydraulic hose 1" (3000 psi all sizes), FT, 6.18
   HH-R12-16, SAE 100R12 4-spiral hydraulic hose 1", FT, 11.76
   HH-R12-20, SAE 100R12 4-spiral hydraulic hose 1-1/4", FT, 16.59
   HH-R13-20, SAE 100R13 6-spiral hydraulic hose 1-1/4", FT, 22.47
   HF-JIC-04-S, JIC 37° female swivel crimp fitting, straight, -04, EA, 1.1
   HF-JIC-04-45, JIC 37° female swivel crimp fitting, 45° elbow, -04, EA, 1.65
   HF-JIC-04-90, JIC 37° female swivel crimp fitting, 90° elbow, -04, EA, 1.76
   HF-NPT-04-M, Male NPT crimp fitting, straight, -04, EA, 0.99
   HF-ORFS-04-S, ORFS female swivel crimp fitting, straight, -04, EA, 1.38
   HF-FER-04, Crimp ferrule for 1- and 2-wire hose -04, EA, 0.37
   HF-JIC-06-S, JIC 37° female swivel crimp fitting, straight, -06, EA, 1.35
   HF-JIC-06-45, JIC 37° female swivel crimp fitting, 45° elbow, -06, EA, 2.03
   HF-JIC-06-90, JIC 37° female swivel crimp fitting, 90° elbow, -06, EA, 2.16
   HF-NPT-06-M, Male NPT crimp fitting, straight, -06, EA, 1.22
   HF-ORFS-06-S, ORFS female swivel crimp fitting, straight, -06, EA, 1.69
   HF-FER-06, Crimp ferrule for 1- and 2-wire hose -06, EA, 0.46
   HF-JIC-08-S, JIC 37° female swivel crimp fitting, straight, -08, EA, 1.95
   HF-JIC-08-45, JIC 37° female swivel crimp fitting, 45° elbow, -08, EA, 2.93
   HF-JIC-08-90, JIC 37° female swivel crimp fitting, 90° elbow, -08, EA, 3.12
   HF-NPT-08-M, Male NPT crimp fitting, straight, -08, EA, 1.76
   HF-ORFS-08-S, ORFS female swivel crimp fitting, straight, -08, EA, 2.44
   HF-FER-08, Crimp ferrule for 1- and 2-wire hose -08, EA, 0.66
   HF-JIC-12-S, JIC 37° female swivel crimp fitting, straight, -12, EA, 3.4
   HF-JIC-12-45, JIC 37° female swivel crimp fitting, 45° elbow, -12, EA, 5.1
   HF-JIC-12-90, JIC 37° female swivel crimp fitting, 90° elbow, -12, EA, 5.44
   HF-NPT-12-M, Male NPT crimp fitting, straight, -12, EA, 3.06
   HF-ORFS-12-S, ORFS female swivel crimp fitting, straight, -12, EA, 4.25
   HF-FER-12, Crimp ferrule for 1- and 2-wire hose -12, EA, 1.16
   HF-JIC-16-S, JIC 37° female swivel crimp fitting, straight, -16, EA, 5.6
   HF-JIC-16-45, JIC 37° female swivel crimp fitting, 45° elbow, -16, EA, 8.4
   HF-JIC-16-90, JIC 37° female swivel crimp fitting, 90° elbow, -16, EA, 8.96
   HF-NPT-16-M, Male NPT crimp fitting, straight, -16, EA, 5.04
   HF-ORFS-16-S, ORFS female swivel crimp fitting, straight, -16, EA, 7.0
   HF-FER-16, Crimp ferrule for 1- and 2-wire hose -16, EA, 1.9
   HF-FER-20, Crimp ferrule for 4-spiral hose -20, EA, 2.1
   HF-CODE61-20, SAE Code 61 flange head crimp fitting -20, EA, 14.3
   SVC-CRIMP, Hose assembly crimp labor, per end, EA, 4.0
   SVC-TEST, Hydrostatic pressure test with certificate, per assembly, EA, 12.0
   
   If a combination isn't on the list, say so instead of guessing. Footer: "Costs from price list dated Aug 2026. Verify before sending."

2. Use it. Then argue with it.

   Quote Marco's kit from capability 4: two 3/8" 2-wire at 42", one at 28" with a 90, one 3/4" at 36" with male NPT. Does the total match? Cramped on your phone? Want a rush fee toggle? Say so. Two or three rounds is normal.

3. Publish it.

   Click the artifact, then Publish. Open the link on your phone. That's the moment.

4. Make it the team's.

   Project "Counter tools": attach your real price list (business plan; yellow data). Instruction: "When asked to update the quote calculator, use the attached price list for costs." Then "update the calculator with October pricing" is one line.

5. Check yourself.

   A working tool on my screen that I built I changed one thing by asking A link on my phone

Or build one of these

Build a web tool for a hose distributor. I paste two CSVs: remnant inventory (remnant_id, sku, branch, length_ft, days_on_hand, bin, unit_cost) and open requirements (quote_no, customer, sku, cut_length_ft, ship_from_branch, need_by).

Match requirements to remnants: same SKU, same branch, remnant length at or above the cut, least waste first, oldest first on ties. A remnant can serve multiple cuts if it has the length. Results table: requirement, matched remnant, waste, dollars recovered (unit_cost times cut length). Totals at the top. Then a liquidation list of unmatched remnants sorted by age. Start with remnant_inventory.csv and open_requirements.csv, and keep paste boxes for next week.

Restate field definitions, units, and deterministic match order before building. Never combine SKUs or branches, accept negative/blank lengths, or allocate more footage than remains. Preserve quote, need-by date, and remnant ID. Show every cut and remaining balance so multi-cut allocation is auditable. Flag duplicate IDs, malformed CSV, zero lengths, missing fields, and no safe match.

Add validation summary, unmatched reason, reset, and export/copy. Process data locally and disclose that it is not persisted. Test exact fit, two cuts from one remnant, least-waste order, age tie, insufficient length, cross-branch rejection, duplicate requirement, and decimals; display pass/fail.

Do not claim inventory is reserved or update an ERP. Footer: “Planning suggestion only; verify physical remnant, bin, SKU, length, and current availability before cutting.” Finish with formulas, assumptions, limitations, and a warehouse/inside-sales acceptance checklist.

Build a one-page quote follow-up board from the attached quotes.csv. Treat its dates and statuses as authoritative; show the file date and let the user set “today” for testing. Columns are overdue for follow-up (open, no completed follow-up, older than 5 calendar days), due this week, waiting on customer, won this month, and lost this month. Define precedence so a quote never appears twice.

Each card shows quote number, customer, amount, rep, branch, quote date, days since quoted, and follow-up state. Totals show count and dollars; overdue sorts by dollars descending. Rep and branch filters must recalculate totals. Flag unknown statuses, missing dates, duplicate quote numbers, invalid amounts, and stale data instead of dropping rows.

“Draft follow-up” creates a subject and three-sentence draft using supported quote facts only. It must never send, change status, or imply current stock or price validity. Add copy and exception-export actions.

Provide tests for the five-day boundary, week/month boundaries, missing follow-up, waiting status, duplicate IDs, and filters, with expected results. Check mobile and keyboard use, formatting, and empty/error states. Explain rules in an About panel. Footer: “Verify quote status, validity, customer contact, and availability before outreach.” A sales manager must reconcile sample cards to quotes.csv and approve the workflow.

Build a rebate tracker from vendor_programs.csv and purchases_by_vendor_month.csv. Identify each file’s data-through date, row grain, vendor key, purchase treatment, tier thresholds/rates, effective period, and retroactive flag. Do not fabricate missing program rules or combine unmatched vendors.

For each vendor show purchases year to date, current tier, next threshold, dollars and percent to it, projected year-end spend at current pace, rebate at projected pace and next tier, and whether retroactive. Display formulas and calendar/program-year assumptions. Handle returns, credits, exclusions, duplicates, partial periods, missing months, and already-exceeded thresholds explicitly. “Within 10%” means distance divided by next threshold.

Add a what-if control to shift $X of eligible Q4 spend from vendor A to B. Validate nonnegative and available shiftable spend; recalculate both vendors and show incremental rebate. Warn that product price, freight, quality, availability, and contract effects are outside the model.

Include source warnings, reset, and export. Test exact thresholds, retroactive versus incremental tiers, partial-year pace, returns, missing tiers, and invalid shifts. Use decimal-safe money and mobile layouts. Footer: “Planning model only; finance and purchasing must verify eligible spend and program terms before moving purchases.” Reconcile totals and provide a human acceptance checklist.

Build a simple web tool for [who uses it and where]. The single job is [decision or task]. Inputs: [what they type, paste, upload, or select, including units]. It should [calculation, match, lookup, or document]. Show [result and exact output structure]. Start with this authentic sample data: [paste or name existing files and dates].

Before coding, restate the user, workflow, source-of-truth fields, formulas, assumptions, and out-of-scope actions. Ask up to five blocking questions one at a time. Do not add login, integrations, databases, or automation to version one. Process data locally unless instructed otherwise and disclose persistence.

Build one clean, accessible, phone-ready page. Validate required fields and units and preserve record IDs. If data is missing, contradictory, unsupported, or unsafe, say “I don’t know” or flag it rather than guessing. Include empty, error, and no-match states. Show calculations and evidence so the result is auditable.

Create six tests with expected outcomes: normal, boundary, missing input, invalid unit/value, duplicate, and unsupported case. Run and display pass/fail. Add only the requested copy/export action and never claim an external record changed. Finish with limitations, data-date footer, and a human checklist naming who verifies formulas, source freshness, operational feasibility, and safety.

Build a web application called Quote Desk for a hose and fittings distributor's inside sales team. The data files in this folder are the seed: price_list.csv, price_levels.csv, margin_floors.csv, customers.csv, quotes.csv.

Users log in (email and password to start; roles rep, manager, admin). A rep creates a quote for a customer picked from the customer master; the customer's price level sets the default discount. Line items are either a stock SKU (searchable) with a quantity, or a hose assembly built from series, size, length in inches, end A, end B, quantity, with cost from hose per foot rounded up to the half foot, two fittings, two ferrules, and crimp labor per end. Every line shows cost, sell, margin %, and warns when margin is below the category floor; below-floor lines need a manager override that's logged. Discontinued or superseded SKUs can't be quoted; the app offers the replacement.

Quotes have status (draft, sent, won, lost, expired), a follow-up date defaulting to quote date plus 5 days, and a lost reason. The quote list highlights anything over 5 days with no activity. Managers see a dashboard: quotes by rep, win rate by response time, average margin, oldest open quotes. Admin uploads a new price list CSV with an effective date and the old costs are kept in a cost history.

Export a quote as PDF and as plain text. SQLite for now. Load the seed data and the existing quotes so it's alive on first run. Plain, fast UI for a counter. README with local run and deploy to Replit or Vercel. CHANGES.md I'll keep updating.

Inspect and document the real columns, dates, keys, and quality issues in all five seed files before implementation; do not fabricate records. Include authentication, password hashing, server-side role checks, validation, immutable overrides, effective-dated costs, and CSV import preview/rollback. Use decimal-safe money. Reject unsupported assembly combinations, duplicates, stale prices, invalid quantities, missing floors, and illegal status transitions with explicit errors. Never expose cost/margin in customer exports.

Write automated known-answer tests for half-foot rounding, assembly components, price levels, margin-floor boundary and override audit, superseded SKUs, role permissions, five-day follow-up, import rollback, and PDF/plain-text output. Include migrations, configuration without secrets, backup/restore notes, and deployment limitations of SQLite. Run lint, type, and tests and report failures. Finish with a manual acceptance checklist for rep, manager, admin, finance, and operations. Treat this as a prototype until security and quote math receive independent review.

Build a small web app for accounts payable that matches vendor invoices to purchase orders and receipts. Seed it from purchase_orders.csv, receipts.csv, and vendor_invoices.csv in this folder, and let me upload new files of the same shape.

For each invoice: match every line to its PO line and receipt by po_no and line_no; compare quantity, unit of measure, unit price, and extended amount; classify exceptions (price variance, short-shipped billed full, over-shipment, UOM mismatch, freight not allowed, surcharge not applicable, duplicate invoice, missing discount) with the dollar impact. Surcharge rule: a tariff surcharge is only valid on lines whose vendor_part_no starts with "N-". Show an invoice list with a status (clean, exceptions, duplicate), an invoice detail page with the line-by-line comparison, an approve / short-pay / query decision per invoice that's logged with who and when, and a vendor scorecard: exception rate and dollars by vendor. Draft the query email from a template. SQLite, plain UI, README, CHANGES.md.

Inspect the three actual files first and document keys, dates, units, currency, discounts, freight, and duplicates. Never force a match when keys conflict; show unmatched and one-to-many cases. Use decimal-safe arithmetic, preserve original and normalized values, and ask me to approve tolerances. Treat approve/short-pay/query as logged proposals, not payment execution. Require authorization, reason, user, timestamp, and immutable history. Draft queries but never send them.

Test exact match, partial receipt, overreceipt, UOM conflict, duplicate invoice, disallowed surcharge, missing discount, freight, rounding boundary, unmatched line, permissions, and upload rollback. Reconcile invoice and exception totals. Validate and escape uploads/exports. Include migrations, backup/configuration, and run/type/test instructions. AP verifies matches, purchasing confirms PO terms, receiving checks receipts, and an authorized approver makes the real payment decision.

Build a public, mobile-ready page where a customer enters a PO number and sees order status from the newest valid nightly CSV (order_no, customer_po, customer_name, status, expected_ship_date, carrier, tracking_no, last_updated). Exact-match customer_po and show all authorized orders. Include a friendly generic not-found result with our verified phone number and an authenticated internal health page showing file timestamp and row count.

Before coding, flag that PO numbers may be guessable and customer_name/order details are sensitive. Propose a safer verifier or authenticated signed-link design and require approval; do not expose data merely because a PO matches. Add rate limits, generic errors, input normalization, secure headers, minimal logging, authorization on health, and no hard-coded password. Validate schema/freshness, ignore temp/partial files, retain last known good data with a stale warning, and sanitize tracking links.

Test authorized and failed lookup, unknown/duplicate PO, stale/malformed/empty export, CSV injection, path traversal, rate limits, mobile/keyboard access, and internal authorization. Provide threat model, environment variables without secrets, deploy/rollback instructions, and CHANGES.md. IT/security reviews the design; operations verifies statuses, dates, and phone number; legal/customer service approves exposed fields before launch.

1. Go back to a chat that went well today. Ask for the instructions.

   Prompt 1 · extract the patternCopy

   Turn the successful task we just completed into reusable instructions for a sharp new hire. No file is needed for this step; use the conversation as evidence. Begin by naming the job, intended user, decision supported, input types, and definition of a good result. Quote or point to the parts of our exchange that establish each rule. Do not treat a one-time preference, correction, or sample value as a universal policy without asking me.
   
   Interview me one question at a time for any missing item that would change the result: source of truth, required fields, definitions, approval authority, edge cases, output audience, or stop conditions. Then write an under-one-page briefing that stands alone outside this chat.
   
   Use sections: when to use; required inputs and freshness; numbered procedure; exact output format; calculations or decision rules; evidence/citation requirements; questions to ask; checks to always run; uncertainty labels; escalation conditions; and what never to do. Separate hard policy, current practice, and suggested improvement. Include at least three tests: normal case, ambiguous/missing evidence, and high-risk exception, with expected behavior.
   
   Never instruct the assistant to invent missing prices, customer facts, technical compatibility, dates, approvals, or system actions. Require a human owner to compare the result to the source and approve anything customer-facing, financial, operational, safety-related, or written back to a system. Finish with a “teach-back” summary of the pattern and ask me to correct it before calling it reusable.

2. Make the Project.

   Left sidebar, Projects, New. Name it for the job ("Quote desk," "Monthly numbers," "Vendor letters"). Paste the instructions into Project instructions. Attach the files that never change (the price list, the price levels, your policy). Start a chat inside it and give it only the new thing: "quote RFQ 2." It already knows how you work.

3. Make the skill.

   Settings, Skills (or ask: "Save this as a skill called Call brief"). A skill is the how-to without the files; your AI assistant applies it in any chat when it fits. Write three today: "Quote desk," "Call brief," "My voice."

   Prompt 2 · write the skillCopy

   Convert the approved briefing from the prior step into a reusable AI skill. No new file is needed. Preserve the job, evidence rules, output, safety boundaries, and human ownership; do not broaden the scope just to make the skill sound capable.
   
   Keep the skill under one printed page and use this exact structure: a short, specific name; one sentence describing when to use it and when not to; required inputs and freshness checks; numbered execution steps; exact output headings/tables; assumptions prohibited; uncertainty and conflict handling; edge cases and stop/escalate conditions; checks before completion; and named human review. Include instructions to cite source file, record/section, and date for material claims and to show formulas for financial or quantity calculations.
   
   Make the skill ask only blocking questions, one at a time. It must distinguish fact, inference, and recommendation; never silently repair data; never claim that an email was sent, a record updated, inventory reserved, or approval received; and never make unsupported technical, legal, credit, pricing, or safety decisions. Add three compact acceptance tests with expected outputs, including one where the correct response is to stop.
   
   After the skill, provide: the exact natural sentence I would type to trigger it; one sentence that should not trigger it; a version/date line; and a five-item owner checklist for testing it on a new example. Ask me to approve or revise the skill after that test before sharing it with a team.

4. Share it. Then run the Monday ritual.

   Invite the team to the Project. Monday: everyone brings one saved prompt to a shared Project called Patterns. Fifteen minutes. That meeting is the whole rollout plan for month one.

5. Check yourself.

   One Project with instructions and a file attached One skill saved Intermediate: one teammate invited

1. Start with the business plan. Then Settings, Connectors.

   On your AI provider's business plan, an admin should decide which connectors are allowed. That is exactly the control IT wants. Turn on one thing, for one person, read-only, for one job.

   1 · Microsoft 365

   Outlook mail and calendar, OneDrive and SharePoint, Teams. Sign in once with your work account. Reading is one permission; sending and writing are separate ones.

   The RFQ sweep, the meeting brief, "find the file," "what did I promise this week."

   2 · The rest

   Google Workspace, Slack, HubSpot or Salesforce, Asana, and a growing catalog. QuickBooks and most ERPs: through an export folder or a small connector (step 4).

   CRM history folded into the call brief; follow-ups logged without typing.

   3 · Your data

   A shared folder your ERP exports to nightly (sales lines, inventory, AR, quotes, open POs). Your AI assistant reads it through the Files connector. That folder is your data lake. No project required.

   "Open the newest file in Exports and run the monthly review."

   4 · The ERP itself

   Rarely a ready connector. A small MCP server over the export or an ODBC/API path; a coding agent can write the first version (the starter is below).

   "How are we doing on Copperline this quarter," asked out loud.

2. Connect Microsoft 365 and run the inbox sweep.

   Settings, Connectors, Microsoft 365, sign in. Then in a new chat:

   Prompt 1 · the RFQ sweep (read-only)Copy

   Use my approved inbox connector in read-only mode for the last five business days, stating the exact start/end dates and mailbox searched. Find customer quote requests, formal RFQs, and informal “can you get me” messages. Search subject and body, inspect full threads, and avoid counting forwards, internal discussion, automated notices, or duplicate threads as separate requests.
   
   For each candidate, report thread/message ID, received date/time, sender/company, requested items and quantities exactly as written, need-by date, attachments, branch/account if supported, whether an employee replied after the request, urgency stated by the customer, and missing information. Mark transcript/OCR or attachment-access gaps. Do not infer SKU, application, stock, price, credit, or urgency from relationship value.
   
   Create an exceptions-first table sorted by confirmed need date, then age. For unanswered items, draft a subject and concise reply in my supplied voice rules, acknowledging the request and asking only blocking clarification questions. Cite the source thread internally. Save as drafts only if the connector supports draft creation and I have approved that action; otherwise present draft text. Never send, forward, change CRM, or promise price/availability.
   
   End with counts searched/found/answered/unanswered, possible misses due to access or search limitations, duplicates merged, and a human review queue. I must open each source thread, verify attachments and reply status, check account/technical facts, and approve each draft before sending.

   Prompt 2 · what did I promiseCopy

   Using approved read-only sent-mail and calendar connectors, review the last seven calendar days and state exact dates, mailbox, and calendars searched. Extract every commitment I personally made to a customer, vendor, or teammate. A commitment requires explicit action or deliverable; distinguish it from a suggestion, request from someone else, meeting invitation, or “I’ll try.”
   
   For each item cite message/event ID, timestamp, exact supporting phrase, recipient, deliverable, stated or inferred due date, conditions, related thread, and confidence. Resolve relative dates from the message timestamp but show the calculation. If no due date exists, say “not stated.” Check later messages in the same thread for evidence of completion, change, cancellation, or acknowledgment; never mark complete merely because time passed.
   
   Output open commitments ordered by due date, overdue first, plus completed/changed items and ambiguous candidates. Draft a neutral one-line nudge for each open item without sending it or inventing status. Flag inaccessible threads, deleted items, delegated commitments, conflicting dates, and calendar events that lack supporting text.
   
   Finish with coverage counts and a human verification step: I open each cited source, confirm ownership/status/due date, and decide whether to send a nudge or update a system. Do not create tasks, send messages, or expose private content beyond the minimum required summary.

   No connector in the room today? Paste customer_emails.txt and run the same prompt. Same result, one step earlier on the ladder.

3. Build the data lake: one folder, nightly.

   Ask your ERP (or IT, or the vendor) for a scheduled export to a SharePoint or OneDrive folder called Exports: sales order lines, inventory on hand, AR aging, quotes, open POs. Same filenames every night. Zero risk: it is the report you already run, landing in a folder. Then:

   Prompt 3 · read the lakeCopy

   Use the approved read-only SharePoint connector. In the Exports folder, identify the newest valid sales_order_lines file and newest valid inventory file by explicit file timestamp; report names, modified dates, data-through dates, row counts, and schema. If either is stale, incomplete, malformed, or on mismatched dates, stop and explain the limitation.
   
   Define top 20 customers by trailing six complete months of revenue unless the files require another supported measure. Exclude the current partial month from the baseline and compare equivalent elapsed days when measuring current month. Show formulas and treatment of returns, credits, transfers, duplicates, zero activity, new customers, and branch changes. For each declining customer report baseline monthly average, current comparable amount, dollar/percent difference, categories driving it, last order date, and confidence.
   
   Determine “usually buy” with a stated frequency rule based on six complete months. Check exact SKU and normal shipping branch against current on-hand. Do not claim out-of-stock when inventory is blank, stale, allocated, or at another branch; label those separately. Cite customer, order rows/filters, inventory SKU/branch, and dates.
   
   Deliver a reconciled table, data-quality exceptions, and the three calls to make today ranked by materiality and actionability, each with evidence and discovery questions rather than a fabricated cause. End with human verification by ERP/data owner, branch inventory owner, and account rep before outreach or transfer decisions.

   Every prompt on this site works on that folder. Match the columns in the file map and the whole guide runs on your business.

4. The ERP direct: a small connector a coding agent writes.

   Coding agent · MCP server over the exportsCopy

   Our ERP exports nightly to an approved local folder: sales_order_lines.csv, inventory_on_hand.csv, ar_aging.csv, and quotes.csv, using the same shapes as the workshop files. Build a small read-only MCP server so an approved desktop AI assistant can query sales for a customer and period, weighted margin by category, open orders, on-hand and days of supply by SKU/branch, AR by customer, and open quotes past follow-up. Add a metadata tool reporting exact file path/name, modified time, data-through date when present, row count, schema validity, and staleness.
   
   Before coding, inspect the actual sample schemas and document keys, date fields, units, formulas, duplicate handling, and source limitations. Ask for the approved folder path and staleness threshold; do not invent them. Define narrow typed tool inputs and bounded outputs. Validate customer/SKU/branch identifiers, date ranges, limits, and requested fields. Use weighted margin formulas, preserve decimal precision, distinguish zero from missing, and return source file/date plus warnings with every result.
   
   Enforce read-only access in design and code: no email, file writes, SQL, shell exposure, arbitrary paths, URL fetching, or record updates. Prevent path traversal and CSV injection, redact unnecessary customer data, cap rows, log tool name/timing/error without sensitive values, and fail closed on stale, malformed, missing, or partially written exports. Document permissions and least-privilege folder access.
   
   Add automated tests for valid query, no match, duplicate keys, malformed/stale/empty file, partial nightly write, date boundary, margin math, access outside the folder, and output limits. Provide simple source code, configuration via environment variables without secrets, setup instructions for the approved assistant, sample tool calls/results, troubleshooting, and uninstall/disable steps. Run tests and report results. IT/security and data owners must review permissions, schemas, calculations, and desktop configuration before use.

5. Turn it on without scaring anyone.

   1. Business plan first. IT controls the connectors.
   2. Read-only for a month. Every prompt ends "find, summarize, draft. Don't send."
   3. One person, one connector, one job.
   4. Approve each action until you're bored. Then let that one action run. Not "always allow everything."
   5. ERP last. The export folder gets you 80% with zero risk.
   I know which connector we'd turn on first, and for whom I know what our Exports folder would contain I have the IT email (on the Trust page) ready to send

1. Write the job in four lines. Every scheduled task is these four.

   Prompt 1 · design the scheduled jobCopy

   Help me design a bounded scheduled job: [7 AM RFQ sweep / first-of-the-month numbers review / Friday vendor letter watch / quote follow-up at 5 days]. Act as an operations and controls designer. Ask one question at a time until you know the approved system/location, cadence and timezone, source date/freshness, business rules, output destination, data owner, reviewer, and actions forbidden.
   
   Start with exactly four lines: READS, including inbox/folder/file and permission scope; DECIDES, including deterministic rules and what it must never decide alone; PRODUCES, including format, destination, naming, and whether outputs are drafts; APPROVAL, naming who reviews before anything leaves or changes a system.
   
   Then write the scheduler-ready task description with trigger, timezone, inputs, schema/freshness validation, deduplication/idempotency key, calculations, edge cases, uncertainty labels, citations, output, audit log, retention, retry limit, timeout, failure alert, stale-data behavior, and kill switch. Default to read-only and drafts. Never authorize sending messages, changing prices/credit/inventory, placing orders, or writing back unless separately approved.
   
   Design a two-week attended test with daily sample checks, expected results, false-positive/false-negative log, source reconciliation, permission review, failure drills, and owner signoff. Define measurable “trust it” thresholds but do not invent acceptable rates; ask me to approve them. Include rollback and escalation. End with a go/no-go checklist for business owner, IT/security, data owner, and output reviewer.

2. Set it up in a desktop AI agent.

   Claude Cowork is one example: an AI assistant on your desktop, working in your files and connected apps for as long as the job takes. In your tool, describe the job (paste the task description from step 1), then Schedule: pick the cadence (weekdays 7 AM, first business day, Fridays at 4). It runs, and it reports back where you told it to. Attended for two weeks means you read the output every morning before anything is acted on.

3. The five jobs that get scheduled first.

   The 7 AM RFQ sweep

   Reads the inbox, drafts quotes from the Project price list, lists the questions, saves drafts. Email + files.

   The Sunday briefing

   Every outside meeting next week gets a researched brief in the Briefs folder. Calendar + search + files.

   The first-of-the-month review

   Export picked up from the Exports folder, four checks, heat map, draft rep notes, deck. Files.

   The vendor letter watch

   Price, surcharge, terms, allocation: decoded, costed against last year's purchases, posted to the purchasing channel. Email + files + Teams.

   The quiet quote follow-up

   Every quote over five days with no reply gets a specific follow-up drafted. Email + the quotes export.

   Yours

   What it reads, what it decides, what it produces, who approves. Three of those, not ten, and it's buildable this month.

4. Simulate it today with the files you have.

   Prompt 2 · the morning brief, by handCopy

   Pretend it is 7:00 AM in the branch’s stated timezone and run the proposed morning job manually. Use only this week’s customer_emails.txt, quotes.csv, inventory_on_hand.csv, and ar_aging.csv. State each file’s name, date/data-through timestamp, row count, and freshness. If a required file is missing, stale, malformed, or has unmatched customer/SKU/branch keys, show an exception and do not hide it.
   
   Produce an under-one-page brief with four ranked sections. RFQs needing a quote today: source message/date, customer, exact need, deadline, missing requirements, and a draft reply that is not sent. Open quotes past five calendar days: quote number/date/value/rep, last follow-up evidence, and one neutral follow-up line. Orders or requests from customers over credit limit: balance, limit, overage calculation, aging context, and “credit review required,” never an approve/hold decision. Requested items out of stock at the customer’s branch: exact SKU, requested quantity, branch on-hand, file date, and alternate-branch information only if supported.
   
   Show rules and calculations, deduplicate messages/quotes, preserve units, cite every item to file and record, and distinguish no record from zero. Cover partial replies, duplicate RFQs, expired quotes, missing credit limits, allocated inventory, and stale counts. Do not promise price, availability, shipment, transfer, or credit.
   
   End with exactly three decisions only I can make, followed by data exceptions and a verification checklist. Inside sales opens source threads and quotes, credit verifies AR, warehouse verifies physical availability, and I approve every draft or action. This manual result becomes the expected baseline for the scheduled test.

   Attach customer_emails.txt, quotes.csv, inventory_on_hand.csv, ar_aging.csv. That is the job. The schedule just removes the attaching.

5. From a scheduled task to an agent.

   An agent is a scheduled job that is allowed to take an action: send the acknowledgment, file the quote, post the summary. Earn it one action at a time: after a month of drafts you barely edit, let it send "quoting by Thursday." Not price notices. Not credit decisions. The rule from the MMG rollout applies here: if a person drives it, it's a pattern you can start Monday; if it has to run on its own connected to your systems, it's a project. Know which one you're asking for.

   One job written in four lines The morning brief produced by hand from the files I know the first action I'd let it take, and when