# The Office Prompt Library All 41 cards from Appendix A of Vibe Coding at the Office, in one file. Copy a card, fill the [BRACKETS], and you are running that chapter's method. Three rules travel with every card. Chapter 3's data rules apply to every slot: nothing confidential goes in, describe instead of paste. The model drafts, you review, you ship. And when a card does not fit your job, change the card. ## How to Read This Appendix Every prompt pattern in this book, one per card, a fill-in-the-blank scaffold with the slots marked in brackets. Copy a card, fill the brackets, and you're running that chapter's method. Three rules travel with every card. Chapter 3's data rules apply to every slot: nothing confidential goes in, describe instead of paste. The model drafts, you review, you ship. And when a card doesn't fit your job, change the card. ## 1. Documents (Chapter 4) The preamble comes first, every time. Audience, outcome, constraints: that is the prompt, and the draft is its output. ### P01. Reader-First Preamble **Use when:** you owe someone a document and the first sentence hasn't been written. > I need [DOCUMENT]. Write it for [AUDIENCE: WHO READS IT, AND IN WHAT SITUATION]. After they finish, [OUTCOME: WHAT YOU WANT TRUE IN ONE SENTENCE]. Constraints: [LENGTH, FORMAT, TONE, MUST-INCLUDES, MUST-AVOIDS]. Source material: [THE FACTS, DESCRIBED]. Draft in one pass; every figure traces to a source I can produce. ### P02. Proposal Responder (answer-first) **Use when:** an RFP or formal request lands with a deadline that now owns your week. > Parse [THE RFP] into a checklist of every question it asks, sub-questions included. Build an outline mirroring its section order exactly. Then draft each section: the direct answer first, then the evidence. No claim I cannot back. Page limit [N], format rules [RULES]. Draft, then redline: mark each section keep, tighten, or cut, with a ten-word reason. Verify: every checklist question has an answer. ### P03. Status-Update Compression **Use when:** the week's update is due and the last one was skimmed or skipped. > My raw week, bulleted: [WHAT HAPPENED]. Write it for [AUDIENCE: SOMEONE WHO SKIMS BEFORE A MEETING]. Under [N] words: what moved, what stalled, what I need and by when. Numbers only where true. ### P04. Plain-Language Summary Recipe **Use when:** a long report needs a front page a boss can forward cold. > Here is [THE REPORT]. Convert it to a one-page executive summary, every number matching the source exactly. Cover: what changed, what it costs, what we decided or must decide, what is still open. Test: a reader who never saw the report agrees with every figure. ### P05. Podium Recipe **Use when:** you are on the podium and you have three points and an open and close. > My brief: [EVENT, AUDIENCE, LENGTH, ONE THING THEY MUST REMEMBER]. Build a scaffold: three points in order, the transitions, the open, the close. Read-aloud pass: flag anything that doesn't sound like a person speaking. Keep each point to [N] lines. If this enters a public record, run Section 8's ladder first. ### P06. Evidence-First Grant Recipe **Use when:** the funder wants impact claims and you will not invent a number. > Draft [SECTION] for [FUNDER]. Every claim maps to a named source in my evidence log: [SOURCES]. Where a claim has no source, leave a [BRACKET]; never a guess. This draft's rule: AI output is a draft to review, not evidence to trust. Attach the source beside every number. ## 2. The Teaching Shelf (Chapter 4) Four classroom recipes, each one drafting first, editing second. ### P07. Standards-First Lesson Recipe **Use when:** you need a first-draft lesson plan you will customize anyway. > Here is the standard: [STANDARD TEXT]. Here is the class: [GRADE, SUBJECT, TIME, WHAT THEY HAVE MET BEFORE]. Draft a lesson plan: objective tied to the standard, opening, activity, check for understanding, close. The plan is a draft for a professional, not a script. Flag spots to customize for [SPECIFIC STUDENTS OR SCHEDULE]. ### P08. Rubric Builder **Use when:** the assignment exists and the grading criteria don't. > Here is the assignment: [ASSIGNMENT TEXT]. Build a rubric: [N] criteria, each with [N] level descriptors written as something observable. The check: two graders land on the same level for the same paper. Descriptors under [N] words. ### P09. Stair-Step Recipe **Use when:** one source text must reach readers at several levels. > Here is the source: [TEXT]. Adapt it to [READING LEVELS, LIKE 3RD AND 7TH GRADE]. Facts survive every level; only the scaffolding changes. For each version: [WHO THIS READER IS, THE TAKEAWAY, VOCABULARY THEY'VE MET]. Include a vocabulary table: words swapped and why. Read the lowest level aloud before you ship it. ### P10. Comment-Bank Recipe **Use when:** twenty grade comments are due in one voice. > Write [N] grade comments in one voice for [CLASS]. Each anchored on one specific, anonymized detail: [TOPICS, LIKE A STUDENT WHO STRUGGLES WITH FRACTIONS]. Each: what the student did, one concrete detail, one next step. No names. Drafts only; I review. ## 3. Spreadsheets (Chapter 5) Four moves, always in order: shape and outcome, options, explain-it-back, edge tests. The original file stays sealed until all four pass. ### P11. Describe-the-Outcome Template **Use when:** a spreadsheet is broken, haunted, or due for a rebuild. > Column headers plus five fake sample rows, same shape: [HEADERS AND ROWS]. The sheet should [PURPOSE IN PLAIN LANGUAGE, LIKE REVENUE TIMES ROLE-BASED RATE, 1.5X ACCELERATOR ABOVE QUOTA]. New requirement: [WHAT CHANGED]. Constraints: [WHAT MUST SURVIVE, LIKE VERIFIABLE IN TEN MINUTES]. What's the cleanest way to structure this? ### P12. Options Question **Use when:** you want choices with tradeoffs, not the first patch. > Give me three ways to [ACHIEVE THE OUTCOME]. For each: what it costs to maintain, what breaks when [A COMMON CHANGE, LIKE A ROW INSERTED] happens, and how I'd tell when it's lying to me. Tradeoffs in plain English. I'll choose. ### P13. Explain-It-Back Demand **Use when:** the proposed solution has to survive a skeptic, and your name is on it. > Explain this design like I'm new to the file: what each piece does, what happens if [A ROW OR COLUMN IS ADDED], and how I'd notice it going wrong. If the explanation doesn't match what I can test, say so. Define jargon on first use. ### P14. Edge-Test Prompt **Use when:** the fix is in place and the weird Tuesday is coming. > Test this on a copy with these edge cases: zero, negative, blank, and [THE BIGGEST VALUE THIS DATA HAS SEEN]. For each: what it did, what it should have done, whether the failure is loud or silent. Report a table. ## 4. Understanding Anything Fast (Chapter 6) Ask the question, get the answer shape you asked for, then check the checkable parts. The count is the part you check. ### P15. Plain-Language Interrogation **Use when:** a pile of numbers needs real questions answered, not "analyze this." > I have a CSV with these columns: [COLUMNS]. Five fake sample rows, same shape: [ROWS]. Answer this question: [THE QUESTION, LIKE HOW MANY TICKETS PER PRODUCT PER MONTH, WITH A TOTAL ROW]. I'll follow up one question at a time. ### P16. Five-Theme Digest **Use when:** leadership keeps asking about the pile and nobody has summarized it. > The pile: [SCRUBBED CONTENT OR DESCRIPTION]. Give me five themes with counts, and one verbatim per theme, copied exactly. Count positions, not topic mentions. The reader must be able to name each theme, its count, and one exact quote. ### P17. Verbatim-Theme Extraction **Use when:** open-ended answers need to become claims you can repeat in a meeting. > Extract the themes from [THE COMMENTS OR VERBATIMS], five or six at most. For each: the count and one exact quote. Keep interpretation separate from counting. Mark themes with fewer than [N] comments as minor. ### P18. Scored Comparison Matrix **Use when:** candidates, vendors, or options need weighing with weights you can defend. > Build a comparison matrix: [OPTIONS] across [CRITERIA]. Weights: [WEIGHTS, LIKE COST 5, MIGRATION RISK 2]. Rate each option on each criterion and show the math. Then the honesty pass: read the table as someone who wants the option I didn't pick. Where does the weighting favor a conclusion I already hold? ## 5. Decks and Reports (Chapter 7) You supply audience, occasion, and one message per slide. The model supplies typography, contrast, and hierarchy. ### P19. Ugly-Deck Rescue Brief **Use when:** the deck you present this week was designed by a committee in 2014. > Audience: [WHO'S IN THE ROOM, WHAT MUST BE TRUE WHEN THEY LEAVE]. Occasion: [MEETING, TIME, DECISION AT THE END]. One message per slide, no walls of text. My outline, one message per slide for the [N] slides I'm keeping: [OUTLINE]. Numbers verbatim from [SOURCE FILE], a source line on every figure: [NUMBERS]. Rebuild clean as [HTML TO PRINT TO PDF, OR NATIVE FILE FOR OUR TOOL]. Keep the logo and colors from [TEMPLATE]. Never invent data. ### P20. One-Pager Conversion **Use when:** the 23-page PDF nobody reads needs a life as one page people tape to the fridge. > Here is [THE LONG DOCUMENT]. Condense to one page, every number verbatim. Top: a decision table ([WHAT TRIGGERS WHAT]). Next: what changed this quarter. Next: the three questions people actually ask, [Q1, Q2, Q3]. Nothing invented, nothing rounded into meaninglessness. ## 6. Email and Communication (Chapter 8) The frame and the dial: cutting and temperature are the two skills. The agent drafts, you send. ### P21. Five-Temperature Rewrite **Use when:** one set of facts needs a register chosen on purpose. > The facts: [FACTS, LIKE MARCH COMMISSIONS LATE, DELIVERY FRIDAY]. The reader: [WHO THEY ARE AND WHAT THEY DECIDE]. Write this at five temperatures: direct, warm, neutral, cautious, formal. Same facts in all five, nothing softened. I'll pick. ### P22. Hard Message **Use when:** the news is bad and the draft keeps drifting into character judgment. > Draft a message to [READER] about [THE NEWS, FACT FIRST]. Name the behavior, never the character: [THE DATE, THE REASON, THE CONSEQUENCE]. Fact line, next-step line, no apology spiral, no "there have been some challenges." Say the date, the reason, what happens next. ### P23. Style-File Test **Use when:** a draft exists and it has to sound like one specific person, including you. > Write this in the voice of the style file. Same facts, same ask: [FACTS]. Then run the test: would [PERSON] open this way, sign this way, and could you read it aloud in their voice without feeling like a fraud? Flag every line that fails, and why. ## 7. Meetings (Chapter 10) A meeting without a record is a rumor with a calendar invite. These cards turn the record into actions with names and dates. ### P24. Transcript to Attributed Actions **Use when:** a meeting ended minutes ago with spoken commitments and no written record. > Here is the raw transcript or my running notes from [MEETING]. Produce four things. Decisions reached, one line each, with who made the call. Action items: owner, exact task, deadline if stated, where tracked. Open questions that need answers before [DATE]. Anything said as "we should probably" that never got an owner. Never invent a deadline; if none was spoken, write "none stated." ### P25. Non-Commitment Surfacing **Use when:** "we should probably" items are aging quietly in the notes. > Here are the notes from [MEETINGS OR THREADS]. List every "we should probably" or "someone should" that never got a name attached. For each: the non-decision, who said it, what deciding it would require. Convert none into tasks; these are decisions needed, with a date. Decide by [DATE], or the non-decision is the decision. ### P26. Monday Chase List **Use when:** it is Monday morning and last week's commitments are about to evaporate. > Here are last week's commitments and today's tracker state: [ITEMS AND TRACKER]. List every item past its deadline or due this week, with its owner. Draft a one-line status request for each. Flag anything marked done that the tracker doesn't confirm. I review every line before anything sends. ## 8. The Inbound Flood (Chapter 11) Sort first, verify second. The ranked digest empties the pile in minutes; the ladder decides what is true. ### P27. Ranked Digest **Use when:** the pile landed and you have ten minutes before it owns your day. > Today's incoming pile, one line per item, nothing confidential: [ITEM LINES]. For each: what it is, what it asks of me, by when, and whether it's a decision I must make or an FYI I can skim. Rank by what needs my judgment soonest, not by sender. Anything that asks nothing and owes nothing goes to the bottom. No stated ask gets "unknown." ### P28. Claim-Extract **Use when:** a document wants your name on it and you need to know what it claims. > List every factual claim in this document as a numbered list. Include numbers, dates, quotes, named sources, anything that asserts something about the world. Do not evaluate. Extract. No confidence scores: confidence is not verification. ### P29. Source-Check **Use when:** claims are extracted and each one needs a home. > For each claim in [THE NUMBERED LIST], find the source it cites, or should have. Three checks per claim: the source exists, the source says what the claim says, the number matches. Quote the evidence and link it. Flag circular citations, a source citing a summary of itself, and stale ones. I'll spot-check three claims myself, two numbers and one quote. ### P30. Citation Validation **Use when:** the citation itself is the claim at stake. > Verify these citations: [CITATIONS]. For each: does it exist, does it support the claim, is it current, primary or a summary of itself? Report a table: claim, source, finding, action: correct, cut, or flag for the author. Never silent. ### P31. Ownership Pre-Flight **Use when:** your name is about to go on a document you did not draft. > I'm about to [SIGN, FORWARD, OR APPROVE] [DOCUMENT], and my name goes on every claim in it. Run the ladder: skim for shape and the ask, extract the claims, source-check each, then a flag-list with four fields: claim, source it should have had, what you found, what I should do. Claims that fail ship corrected or flagged. Never silent. ## 9. The Advisory Board (Chapter 12) One persona per request, one job each, and the human decides last. ### P32. Skeptic **Use when:** you need the assumption this thing cannot survive. > You are the Skeptic. Your one job: find the assumption this [MEMO OR DECISION] cannot survive. Read it, then attack the reasoning, not the writing. Three strongest objections, weakest to strongest. Answer one at a time. ### P33. Audience **Use when:** the document goes to people with no context and no patience. > You are the Audience. Read this the way [A FRONT-LINE EMPLOYEE OR A STRANGER] would: no context, no patience, a rumor heard yesterday. Can they tell what this means, whether their job changes, what they're supposed to do? Flag lines that assume knowledge the reader won't have. ### P34. Executive **Use when:** the ask, the cost, and the risk must survive one minute in a room. > You are the Executive. You have one minute, a budget, and a reputation. Read [THE MEMO]: name the ask, the cost, and what breaks. Kill vagueness by force: flag anything that can't survive a follow-up question. ### P35. Newcomer **Use when:** the artifact assumes the reader already knows the "we." > You are the Newcomer. You joined last week. Read [THE ARTIFACT] and tell me what it assumes I already know: the jargon, the in-jokes, the "we" I was never part of. List every term that needs defining. ### P36. Role-Play Setup **Use when:** a hard conversation is coming and your lines need rehearsal, not scripting. > Goal first: when this conversation ends, I want [SPECIFIC OUTCOME, NOT "WIN"]. The person: [ROLE, ANONYMIZED]. Play them at three stances: cooperative, defensive, emotional. Push back in character; I answer. Then extract the keeper phrases: the lines that held across all three stances, the ones that died, what the other chair heard. Behavioral observations only, never personality judgments. ### P37. Negotiation Rehearsal **Use when:** you are about to ask for something you have been avoiding asking for. > I make this ask: [THE ASK]. You play [THE OTHER SIDE] at three hostility levels: warm, firm, cold. Run it three times; I answer in character. Then swap chairs once: you play me, I play them. Extract the keepers: lines that survive every run. Then I'll say them aloud; flag anything I can't say without wincing. Anonymize everything: no real names, no real numbers. ### P38. Three-Readings Decode **Use when:** a short message arrives and your reply commits you. > Here is the message: [MESSAGE]. What was actually asked? Three readings: one charitable, one literal, one guarded. For each: what it changes about my reply. Then draft one reply that serves all three readings. ## 10. The Recurring Report Factory (Chapter 9) Describe the chain, watch three runs, then let it run. Everything it produces arrives as a draft, and review is the shipping gate. ### P39. Pipeline Description **Use when:** a report you regenerate by hand more than twice needs to build itself. > Every [FREQUENCY] a file lands in [FOLDER] with [WHAT IS IN IT]. Compute: [THE CHANGES AND DELTAS, VERSUS THE PRIOR PERIOD]. Summarize: what moved and why it matters, in my voice (style file attached). Flag anything that looks off instead of smoothing it over. Output: [ARTIFACT SHAPE, LENGTH, FORMAT]. Draft only; I review and ship. ### P40. Drift Check **Use when:** the factory has run for a while and the world has changed around it. > Compare the last three outputs side by side: [OUTPUTS]. Check: same voice, numbers match their sources, shape matches the pipeline description. List every drift: renamed fields, new columns, changed formats, stale phrasing. For each: what changed, and what the prompt needs to know. ### P41. Monitor Setup **Use when:** the factory runs without you and silence is the failure mode. > Job: [DESCRIBE THE RUN]. Three monitors. A failure alert if the run fails or the source never lands. A last-run timestamp I can glance at. A monthly quality audit comparing this month's draft to last month's. The check that the check ran is mine. Nothing ships without my review. ## 11. Computer and Browser Use (Chapter 13) The screen-level chores don't get cards here. They run terse in Appendix G: trigger, steps, guardrails, verify, then the judgment handed back to you. Keep the Chapter 13 decision rule in your head: API or file, script it; web app, browser agent; everything else a human can click, a computer-use agent. Run the worth-it test first: minutes times how often the chore recurs; if setup won't pay for itself in a month, leave the chore alone.