You have a document to write that someone will decide on. A proposal to a client, a quarterly report for a board, a memo recommending a vendor, an application that has to carry your case in two pages. The deadline is real, the reader is busy, and the document will be read by people who will ask where a number came from. You have the facts in six places and no draft.
This is the format where the assistant's two biggest defaults do the most harm and the most good at once. The harm: it will fill a table with plausible figures, recommend a phased approach, and declare that stakeholders have been consulted, all in a register that sounds authoritative. The good: it is very fast at structure, at turning a page of notes into sections with the right headings, and at listing every claim in a draft so you can check them.
Vibe writing for a professional document keeps a hard line. The decisions and the numbers are yours. The structure is built from the reader's decision, never from a template. The executive summary is written last, by you. Every table traces to a spreadsheet you own. And the document goes through whatever review chain your organization uses, with its AI assistance disclosed where the policy asks. The assistant proposes everything else.
What changes for this format
The method is the same everywhere. These are the places where this format bends it.
Decisions and numbers belong to you
A report, proposal, or memo exists to carry a recommendation and the evidence for it. Both have to come from you before the chat opens. Ask the model to recommend and it will recommend the option that offends no one, usually a phased rollout. Leave a gap in the numbers and it will fill the table with figures that look right and sum to something. The brief states the decision in one sentence, puts every number in the raw material block with its source, and instructs the assistant to write [VERIFY] wherever a figure is missing rather than estimate. If a number in the draft is not in your material, it is invented, no matter how reasonable it looks.
Structure from the reader's decision
The model's default is Background, Analysis, Options, Recommendation, Next Steps, which follows the order in which the writer thought rather than the order in which the reader decides. Start from the person who signs: what do they need to know to say yes, what will they object to, what happens if they say nothing. The structure that follows is usually Recommendation, cost, the one metric that justifies it, risk, what the alternative costs, appendix. Give the assistant the reader's decision and objections as direction and ask for structure first, prose second.
Executive summary last
The model writes the executive summary first, because it is at the top, and what it writes is a summary of the outline. The executive summary is the only part most readers will finish, so it has to be written from the finished body, by you, in one paragraph: the recommendation, the cost, the deadline for the decision, the main risk. Ask the assistant to leave that section as a placeholder until the end. If you want help, give it the finished body and ask for a one-paragraph summary to react against, then write your own.
Tables and claims
The model will produce tables. It will also produce totals that do not sum, percentages computed on the wrong base, and a column of figures it has quietly rounded. Build every table in your spreadsheet and paste it as source material. Ask the assistant to describe what the table shows, never to compute from it. Then check every number in the prose against the table yourself, including the ones that look obvious, because the obvious ones are where a transposed digit hides. A claim in a professional document is a promise, and the reader will hold you to it, not the tool.
Review chains and disclosure at work
Documents like these go through people: a manager, finance, legal, a client's procurement team. Sentences that make commitments, we will deliver by the end of Q2, need a human owner who agreed to them. Mark AI-drafted sections as drafts in the file during review so reviewers know what to read hard. Check your organization's policy on AI assistance. Many now ask for disclosure on certain document types, and some prohibit pasting internal data into external tools at all. Where the policy is silent, tell your reviewer what the assistant did. It costs nothing and it means the review is reading the right things closely.
Applications and cover letters are the same discipline
A cover letter or a grant application is a professional document with a reader who decides, and the model's default is a list of the adjectives on your resume. The Cover Letter mega prompt handles the format. The principle is the one above. The raw material is the posting and your actual examples, with numbers: the project, what you did, what changed. The decision is what you are asking for. Nothing in the letter can be true of a different applicant, and nothing in it can be something you did not do.
The process
- 01You
Write the decision sentence and the reader's objections
One sentence: I recommend X, at a cost of Y, and I need a decision by Z. Then the objections the reader will raise, in their voice, however many there are. If you cannot write the sentence, you are not ready to draft: you are still deciding, and the assistant will not decide for you in any way you would want to sign. These lines go into the accepted-decisions ledger as protected.
- 02You
Assemble the source note
Every number with its source: the spreadsheet, the quote from the vendor, the dated report, the trial results. Every constraint: budget, policy, deadline, what was already promised. Every table, built in your own tool. This is the raw material block, and the brief makes it the only source of facts. Replace confidential figures with placeholders if your policy requires it, and fill them in by hand at the end.
- 03AI proposes
Get a structure proposal before any prose
Use the Professional Document mega prompt. Give it the decision sentence, the reader, the objections, and the source note. Ask for structure only: section headings, what each section must establish, which objection each one answers, and where the tables go. Reject a Background-first outline. The reader does not need background to decide. They need the recommendation and the one number that justifies it.
Send this Do not draft yet. The decision this document must carry: [DECISION SENTENCE]. The reader who signs: [ROLE]. Their likely objections: [OBJECTIONS]. Propose a structure in which the recommendation comes first, each section answers one objection, and the executive summary is a placeholder I will write last. For each section, one line on what it must establish and which source in my material it draws on. Then wait.
- 04AI proposes
Draft section by section, numbers only from source
One section per turn, in the accepted structure. For any section with a table, paste the table and ask for prose that describes it rather than prose that calculates from it. Watch for the model's commitments: we will, the team has, stakeholders agree. Each of those needs an owner, and if nobody has agreed to it, it becomes [VERIFY] or is cut.
- 05You
Fact-check pass and table audit
Run the Fact-Check Pass so the assistant lists every number, every claim, every commitment, and every place it could not trace a figure to your source note. Then audit with the spreadsheet open: every number in the prose matches the table, every total sums, every percentage has the right base. The assistant cannot verify any of this. If its list says all figures confirmed, the brief has failed, and you check harder.
Send this Do not rewrite anything. List every number, percentage, date, named person, and commitment in the draft, one per line, with the exact place in my source material it came from. Mark any you cannot trace as UNTRACED. Mark any sentence that commits us to an action or a date as COMMITMENT. Do not claim to have verified anything; you cannot. Then wait.
- 06You
Write the executive summary by hand
With the body final, write one paragraph: the recommendation, the cost, the decision deadline, the main risk and what you propose to do about it. Write it for the reader who will read nothing else. If you want a foil, ask the assistant for its one-paragraph version of the finished body and notice what it leaves out. That is usually the risk, because the model prefers a confident close.
- 07You
Review chain, disclosure, final edit
Mark which sections were AI-drafted, send the document through your organization's review, and tell reviewers what the assistant did if the policy asks or is silent. Run the Final Human Edit for the sentences that sound like a template: significantly enhance, position us for growth, comprehensive evaluation. Fill in any placeholders by hand. Then read it once as the person who signs, and ask whether every number in it is one you could defend in the meeting.
Toolkit
The prompts that fit this format, and when to reach for each.
- Mega promptProfessional Document mega promptThe session set-up for reports, proposals, and memos. It carries the decision sentence, the reader's objections, and the numbers-only-from-source rule.
- Mega promptCover Letter mega promptFor applications and cover letters. Same discipline, with the posting and your own examples as the only raw material.
- Stage promptFact-Check PassAfter the full draft, with the spreadsheet open. It lists the numbers and commitments, and you trace each one.
- Stage promptFinal Human EditThe last pass before review. It catches template sentences, leftover placeholders, and the confident close the model adds when nobody asked.
Before and after
A paragraph the way an assistant tends to return it, and the same paragraph after a writer took it back.
This memo outlines our recommendation to transition to a new customer support platform. After a comprehensive evaluation of several leading solutions, we believe that migrating will significantly enhance our operational efficiency, improve customer satisfaction, and reduce long-term costs. The proposed platform offers robust automation capabilities, seamless integrations, and an intuitive user interface. We recommend a phased rollout beginning in Q1, with full adoption by the end of Q2. Key stakeholders have been consulted, and initial feedback has been positive. We are confident that this transition will position the team for future growth.
I recommend we move support ticketing to the second vendor we trialed, at $14,400 a year, and I need a decision by October 30, before the current contract renews in November. The current tool costs $9,600 a year. The extra $4,800 buys the one thing we tested for: over the four-week trial, median first reply time was 41 minutes on the new tool against 2 hours 10 minutes on the current one, same three agents. Migration is two days of my time plus the vendor's import. The risk I cannot yet rule out is the 140 open tickets; the import has lost ticket history for other customers [VERIFY: confirm with two of the vendor's references].
The decision and the numbers are now the writer's: the vendor, $14,400 against $9,600, 41 minutes against 2 hours 10, the 140 tickets, the October 30 date, all from her source note, where the AI draft had significantly enhance and reduce long-term costs with nothing behind them. The structure follows the reader's decision: recommendation and deadline, cost difference, the one metric that justified it, the risk, instead of comprehensive evaluation and phased rollout. And the unverified claim is marked rather than smoothed. The [VERIFY] tag stays in the draft until she has called the references, and the sentence about stakeholders being consulted, which was not true, is gone along with robust and seamless.
Pitfalls
A table in the draft has figures you did not provide.
They are invented. Delete the table, paste your own from the spreadsheet, and ask only for prose that describes it. Add numbers only from my material, otherwise [VERIFY] to the brief.
The recommendation is a phased approach with a steering committee.
The model had no decision to carry. Write the decision sentence yourself, with the cost and the deadline, make it protected in the ledger, and brief the draft from it.
The totals do not sum, or a percentage is on the wrong base.
Never let the assistant compute. Build tables in your own tool, audit every number in the prose against them, and treat the obvious ones with the most suspicion.
The draft commits the team to dates nobody agreed to.
Run the Fact-Check Pass with a COMMITMENT marker, and give every commitment a human owner or cut it. A sentence that promises on someone else's behalf is a claim you cannot defend.
You are not sure whether to tell your manager the draft was AI-assisted.
Read the policy. If it asks, disclose. If it is silent, tell your reviewer which sections the assistant drafted so they read those closely. There is no version of this where hiding it helps you later.
The cover letter is a list of the adjectives on your resume.
Use the Cover Letter mega prompt with the posting and two real examples with numbers as the only material. Cut every sentence that could be true of another applicant.
Questions
Can an AI assistant write a business proposal for me?
It can draft one quickly from your decision, your numbers, and your reader's objections, and the result will need an hour of judging rather than a day of writing. It cannot supply the decision or the numbers, and when it tries, the proposal looks confident and is wrong. Treat it as a fast structural drafter working from material you own.
How do I keep the numbers in a report accurate?
Every number lives in your source note with where it came from, tables are built in your own spreadsheet and pasted as source, and the assistant is told to describe tables rather than compute from them. After the draft, run a pass that lists every figure and its origin, and check each one with the spreadsheet open. The assistant can produce the list, but only you can do the checking.
Do I have to tell my manager I used AI for a memo?
Read your organization's policy first. Many now specify. Where it is silent, tell your reviewer which sections were AI-drafted so they know what to read hard. Disclosure in a review chain is less about ethics than about getting a useful review, and it protects you if a figure turns out wrong later.
Is the method different for a cover letter or an application?
The discipline is the same with the stakes concentrated: the reader decides, the facts are yours, and nothing can be true of another applicant. Use the Cover Letter mega prompt with the posting and two or three real examples with numbers as the only raw material. Cut every adjective the model adds about you. They are the lines a hiring manager has read a thousand times.
What should never be pasted into the chat for a work document?
Anything your organization's policy names, and in its absence: client data, employee information, unreleased financials, legal matters, and anything under NDA. Use placeholders and fill them in by hand in the final file. If a document cannot be drafted with placeholders, it is a document to write without a third-party tool.