Draft a Report or Proposal with AI (Where the Thinking Stays Yours)
Turn your notes, a task export and your template into a monthly status report.
Monthly reports, weekly updates, project proposals: AI can turn a pile of notes into something that looks finished in under a minute. That’s exactly the risk. A report that looks finished but says nothing specific, or says something you didn’t decide, costs your readers time and costs you credibility. In the next 30 minutes you’ll draft a real report where you make the calls, the AI does the writing, and every sentence can be traced back to your material.
Remember the intern: smart, fast, first day on the job. They write a clean report from whatever you hand them. What they can’t do is judge whether the project is really on track, which risk your director cares about, or what you should ask for. If you don’t say, they’ll fill the gap with confident, general sentences that could be in anyone’s report.
What you’ll get
| The result | A first draft of a one-to-two page report in your company’s format, built only from your material, with your conclusions on top and open items marked |
| Level | Level 2 (an assistant that reads your files) |
| Loop step you’ll practice | “Give context” and “give instructions”: your material, your template, and above all your judgment, before any drafting. Plus “you check”: every claim traced to a source |
| Time | 30 minutes, plus 10 minutes to save your template as a Skill |
What you need
- Claude, any plan. Other tools work too; see the end.
- Your material. We’ll use this throughout: a monthly status report on an internal software rollout, for your director. You’ll need three things:
- Your notes from the month (meeting notes, emails you forwarded to yourself, a running to-do list)
- A data export, such as a task list with due dates and status
- Your company’s report template, or last month’s report as an example
- Your own view of how it went. That’s Step 2, and the most important input.
Check what you’re allowed to upload. Internal project notes may be fine in your company’s approved tool and not in a personal account.
The three rules
| Rule | What you actually do | Why |
|---|---|---|
| Your conclusions first | Before any drafting, write down the overall status, the biggest risk, and what you need from the reader | Research found AI makes work look more convincing whether or not the conclusion is right (see “Why checking matters”). The judgment has to come from you |
| Only from your material | Tell it to use only what you uploaded, and to mark gaps as [TBD] instead of filling them | Otherwise the intern fills gaps with plausible, general sentences, which is how polished-but-empty reports happen |
| Read it as your reader | Make it list where each claim comes from, cut generic sentences, and fix every [TBD] yourself | Your name is on it. Readers judge senders of empty AI reports harshly |
Part 1: One report, drafted properly (30 minutes)
Step 1. Gather the material and see what’s missing (4 minutes)
Put your notes, the data export and the template in one place. Skim them and write down anything you know is missing: “no update from the training team yet”, “budget figure for October not final.” These will become [TBD]s, and that’s fine. A draft that shows its gaps is more useful than one that hides them.
Step 2. Write your judgment in three lines (4 minutes)
On paper, before the AI sees anything:
- Overall status: “Slightly behind: two of five teams are live, three planned for next month.”
- Biggest risk: “The data migration for the finance team, which depends on a vendor fix.”
- What I need from the reader: “A decision on whether to delay the finance team by two weeks.”
These three lines are what your director actually reads. Everything else supports them.
Step 3. Brief the intern and upload the material (6 minutes)
Open a new chat in Claude, upload your notes, export and template, and paste:
Draft this month's status report using the attached template.
Reader: my director. She reads it in 3 minutes and needs to make one decision.
My conclusions (use these exactly; do not change or soften them):
- Overall status: [your line]
- Biggest risk: [your line]
- What I need: [your line]
Rules:
- Use only facts from the attached notes and data. Don't add progress, reasons, benefits or plans that aren't there.
- Where the template asks for something my material doesn't cover, write [TBD: what's missing].
- Put my three conclusions at the top.
- No generic phrases like "significant progress" or "strong collaboration". Use specific facts or leave it out.
- Maximum 2 pages.
The line “use these exactly” keeps the judgment yours. The [TBD] rule keeps the intern from inventing.
Step 4. Interrogate the draft (5 minutes)
Now check your own draft:
1. For every factual statement, tell me which file and which part it came from.
2. List any sentence that isn't directly supported by my material.
3. List any sentence that could appear in almost any project report (too general to be useful).
Delete or fix everything in points 2 and 3.
Step 5. Fill the gaps and check the numbers yourself (6 minutes)
Go through every [TBD]. Fill it with real information (send a quick message to whoever knows) or replace it with an honest “not yet available, expected [date].” Then check every number against your data export. Counts, dates and percentages are where copying errors hide.
Step 6. Cut, read aloud, and own it (5 minutes)
Cut this draft by about 20% without removing any fact, number, date or my three conclusions.
Then read the final version out loud, once, as if presenting it. Change anything you wouldn’t say. If you’d struggle to answer a question about a sentence, either find out the answer or cut the sentence.
Part 2: Save your report format as a Skill (10 minutes, once)
In the framework, “write the weekly report in our company template” is the classic example of a Skill: a manual on the intern’s shelf that Claude opens by itself whenever you ask for that kind of report.
- Start from a report that worked. Only after Part 1 produced a draft you were happy with.
- Ask Claude to build the Skill. In the same chat, or a new one with your template attached:
Turn the way we just wrote this report into a Skill called "monthly-status-report". Include:
- The template structure (attached)
- That I always give my three conclusions first, and you use them exactly
- That you use only facts from the material I upload, and mark gaps as [TBD: what's missing]
- The "any report" test: no sentence without a name, number, date or decision
- After drafting, list the source of every factual statement and any unsupported or generic sentence
Use it whenever I ask for a monthly status report.
- Save it and check it’s on. Save the Skill as Claude instructs, then check Settings > Capabilities > Skills.
- Next month: upload the new material, write your three lines, and say “monthly status report.” Still do Steps 4 to 6 yourself.
The same works for any document with a fixed format: weekly updates, project proposals, incident write-ups.
Why checking matters
The danger with AI-drafted reports isn’t bad grammar. It’s that they look good regardless of what’s in them. Researchers have measured both sides of that:
- AI makes wrong conclusions more convincing. In a 2023 experiment with about 750 Boston Consulting Group consultants, those using GPT-4 did about 40% better on creative product tasks. But on a business problem that required careful reading of financial data and interview notes, they did about 23% worse than consultants without AI. GPT-4 was persuasive even when its recommendation was wrong, and many consultants went along with it (BCG authors in Fortune, Oct 2023). That’s why Step 2 has you write your conclusions before the AI writes anything.
- Readers notice empty AI work, and judge the sender. In a 2025 survey of 1,150 US desk workers, 40% said they’d received AI-generated work in the past month that looked polished but lacked substance (the researchers call it “workslop”). Each case took nearly two hours to deal with, and about half of recipients saw the sender as less creative, capable and reliable (Axios, on BetterUp Labs and Stanford research in Harvard Business Review, Sept 2025).
The consulting study used a 2023 model, and the survey is self-reported. But together they explain the method: your judgment on top, only your facts underneath, and no sentence that could be in anyone’s report.
Checklist
Before you send the report:
- The overall status, biggest risk and request at the top are mine, word for word
- Every factual statement traces back to my notes or data
- Every number matches the data export
- No [TBD] is left, or it honestly says what’s missing and when it’s expected
- No sentence fails the “any report” test
- I read it aloud and could answer a question about every sentence
- The material I uploaded was OK to put in this tool
What you just learned
On the surface you wrote a report. Underneath, you practiced the first two steps of the loop in their most important form: giving context (your material) and giving instructions (your conclusions and rules), before the AI drafted a word. Then you practiced checking by tracing every claim back to its source. The AI wrote; you decided.
The same method works for: a project proposal, a business case for a purchase, an incident report, a quarterly review, and a handover document when you change roles. Change the material and the three lines in Step 2; keep “only from my material” and the “any report” test.
Using a different tool?
The prompts work anywhere. Only the buttons change.
| Tool | For Part 1 | For Part 2 |
|---|---|---|
| Claude | Upload files in a chat; ask for a Word file if you want one | A Skill, managed in Settings > Capabilities > Skills |
| ChatGPT | Upload files in a chat | Project instructions with the template in the project’s files |
| Gemini | Upload files in a chat, or use Gemini in Google Docs with your Drive files | A Gem with the template as a knowledge file |
| Microsoft 365 Copilot only | In Word, use Copilot to draft from your files | Keep the Step 3 prompt and your template in one document and reuse it each month |
Features and plan names change often. This table was checked in October 2026; the ChatGPT, Gemini in Docs and Copilot in Word rows still need checking against their help pages.
Sources
- BCG’s experiment with GPT-4 and 750 consultants (BCG authors in Fortune, Oct 2023), on Dell’Acqua et al., “Navigating the Jagged Technological Frontier”
- AI “workslop” sabotages productivity, study finds (Axios, Sept 2025)
- How to create a skill with Claude through conversation (Claude Academy)
Get new tutorials by email
One email when a new tutorial is out. No spam, unsubscribe any time.
Almost done: check your inbox and click the link in the email to confirm.
That didn't work. Check the address and try again.
I only use your address to send new tutorials. Privacy