A good prompt does not make an AI tool correct. It makes the tool's mistakes easier to see. That is the idea behind this collection of AI prompts for project managers: each one is written so that its output can be checked against something you already hold, such as a log, a schedule or the minutes of a meeting.
This guide covers the anatomy of a reliable prompt, ten prompts for everyday project work, a worked example of verification, and a list of what never to paste into a tool. It complements our AI governance policy template: the policy says what is allowed, and the prompts show how to work within it.
The anatomy of a prompt you can verify
Five parts cover most project prompts.
- Role. Tell the tool what perspective to take, such as an experienced project manager reviewing a plan.
- Evidence. Supply the material it must use, and say that it should use nothing else.
- Task. State exactly what you want, in one or two sentences.
- Rules. Set constraints: tone, length, what to do when information is missing, and a requirement to flag uncertainty.
- Format. Specify the structure, such as a table with named columns, so that checking is quick.
Two rules matter most. Restrict the tool to the evidence you supply, and tell it what to do when it does not know, for example "write NOT STATED rather than guessing". Our post on AI in project controls explains why these two instructions remove so many errors.
Ten AI prompts for project managers, each with its check
In the prompts, replace the capitalised placeholders with your own content. Use only data that your organisation permits in the tool you are using.
1. Draft a weekly status report
Act as a project manager writing for a sponsor. Using only the log below, draft a status report with: summary (three sentences), progress against milestones, top three risks, decisions needed. If a fact is missing, write NOT STATED. LOG: PASTE ANONYMISED LOGCheck: compare every milestone date and status with the log. Highlight any sentence you cannot trace.
2. Turn meeting notes into minutes
From the notes below, produce minutes with: decisions made, actions (owner, due date), open questions. Use only information in the notes. If an owner or date is missing, write NOT STATED. NOTES: PASTE NOTESCheck: read the actions out in the next meeting. Anything that attendees do not recognise is a likely invention.
3. Sharpen a risk description
Rewrite each risk below in the form: cause, event, effect. Do not change the meaning, add facts or suggest new risks. RISKS: PASTE RISK TEXTCheck: the meaning must be unchanged. Ask the risk owner to confirm the rewritten wording.
4. Cross-check the risk log against the schedule
Below are a risk log and a list of schedule milestones. List risks that mention activities or dates not in the milestone list, and milestones with no related risk. Do not invent items. Output as two tables.Check: verify each flagged item manually. Treat the output as a to-do list for review, not as a finding.
5. Draft a stakeholder update with the right tone
Draft a short email to a senior stakeholder about a two-week delay. Facts: PASTE FACTS. Tone: factual, calm, no blame. Include: what happened, effect on the date, what we are doing, what we need from them. Do not add facts.Check: confirm every fact against your own record, and read the email as the recipient would.
6. Generate clarifying questions for a vague request
A stakeholder asked: PASTE REQUEST. List ten questions I should ask to clarify scope, acceptance criteria, constraints and dependencies. Group them by topic.Check: discard irrelevant questions. This prompt creates no facts, so the risk is low.
7. Summarise lessons learned
From the lessons-learned notes below, group the points into themes. For each theme give a one-sentence recommendation for future projects. Quote the note number supporting each theme. Notes: PASTE NOTESCheck: follow each quoted note number back to the source and confirm it supports the theme.
8. Prepare for a change request review
Here is a change request. List the questions a review board should ask about cost, schedule, risk, quality and benefits. Mark each question as ANSWERED IN THE REQUEST or NOT ANSWERED. Request: PASTE REQUESTCheck: read the request yourself for the questions marked answered.
9. Run a pre-mortem
Assume this project failed. Based on the plan summary below, give ten plausible reasons, ranked by how plausible they are given the plan. For each, name one early warning sign I could monitor. Plan: PASTE SUMMARYCheck: the ideas are prompts for discussion. Bring them to the team and add only those they recognise.
10. Draft a meeting agenda
Create a 45-minute steering meeting agenda from the items below. Allocate time to each item, put decisions first, and note the pre-reading for each. Items: PASTE ITEMSCheck: confirm that the time totals match and that decisions are not buried.
A worked example: verification in action
This example uses fictional data. A project manager runs prompt 1 on an anonymised log with these facts: milestone M1 achieved on time, M2 achieved nine days late, M3 not yet due. The draft says: "Two milestones were achieved on schedule and the third is progressing well."
| Draft statement | Source check | Action |
|---|---|---|
| Two milestones achieved on schedule | M2 was nine days late | Corrected to one on time, one late |
| Third is progressing well | Log has no progress note for M3 | Replaced with NOT STATED; owner asked for a status |
| No risks reported | Log lists two open risks | Added with owners |
The prompt rules were good, and the tool still produced an optimistic summary. The check took five minutes and prevented a misleading report. Verification is not optional, even when the prompt is carefully written.
What never to paste into a tool
Unless your organisation has approved a specific tool for that class of data, avoid entering:
- personal data about colleagues, suppliers or customers;
- confidential contract terms, rates and commercial positions;
- unreleased financial results or forecasts;
- security details, credentials or access information;
- legally privileged advice;
- anything covered by a confidentiality agreement that does not allow it.
Practical habits help. Anonymise names, strip rates and substitute generic labels before pasting. Ask your information security or data protection contact when in doubt. The data classes in the policy template give a simple way to decide.
Testing a prompt before it joins the library
A prompt earns its place in a shared library by passing a test, not by sounding good. The simplest test is to run it on a case where you already know the right answer: last month's log, for which the signed-off report exists, or a set of minutes that attendees have already approved. Compare the output line by line and record what the tool got wrong, such as inventing an owner, softening a late milestone or dropping a risk. Then tighten the prompt against those specific failures and run it again. Keep the test case and the failure notes beside the prompt, so that when a tool or model changes you can rerun the same test in minutes and see whether the wording still holds.
Habits that make prompting reliable
- Keep a shared prompt library so that the team uses tested wording.
- Iterate in steps. Ask for structure first, then content.
- Ask for the working. For any number, request the calculation, then recompute it yourself.
- Ask for counter-arguments. A tool that only agrees with you adds little.
- Log AI use as required by your policy, and disclose AI assistance where it matters.
- Never forward AI output unreviewed to a sponsor, client or regulator.
Where this fits in your development
Using AI well is becoming a normal part of project leadership. The PML-AI credential from PCI AI (opens in a new tab) covers AI-enabled project management alongside governance, planning, execution and agile and hybrid delivery, and its exams are scenario-based; see our PML-AI certification guide. For reporting and measurement prompts, see our KPI dashboard guide, and for the split of responsibilities between roles, project controls vs project management.
Tools that help / Learn it properly
To build prompting skill systematically, Optimize All offers free courses: Prompt Engineering Foundations for the structure of good prompts, Mastering Claude and Mastering ChatGPT for working with those assistants day to day. For the leadership context, see Project Management Leadership with AI, and for the credential detail see our PCI AI partner page.
Frequently asked questions
Can I use these prompts in any AI assistant?
Yes, in principle, because they rely on structure rather than a specific product. Always follow your organisation's rules on which tools may be used with which data.
How do I stop the tool inventing facts?
Restrict it to supplied evidence, tell it to write NOT STATED when information is missing, and verify the output against source. No wording removes the need to check.
Should I tell stakeholders when AI helped?
Follow your organisation's policy. A short disclosure line on AI-assisted material builds trust and shows that the work was reviewed.
Which prompt should I start with?
Start with one where errors are easy to spot and low in consequence, such as meeting minutes or agenda drafts, then build up to reports and analysis.
Optimize All is the official marketing partner of PCI AI and Certuvo.