AI prompts for product managers that show their evidence

"Summarise these interviews" is the AI equivalent of a ticket that says "improve onboarding". You'll get something back. It will look finished. And it will reflect the model's guesses about what you meant, which is the part nobody reviews.

Anthropic's prompting guide has a test that suits product people well: treat Claude as "a brilliant but new employee who lacks context", and show your prompt to a colleague with minimal context. "If they'd be confused, Claude will be too." OpenAI's guidance for its recent models makes the same point from the other side: describe "the expected outcome, success criteria, allowed side effects, evidence rules, and output shape", and tell the model not to invent "specific names, first-party data claims, metrics" (OpenAI, GPT-5.5 guide).

Every prompt on this page follows three habits that come from that guidance:

• Say what counts as evidence. A quote, an issue link, a query result. Not "insights".
• Give it an out. "If the material doesn't say, write OPEN QUESTION." Anthropic says permission to admit uncertainty "can drastically reduce false information".
• Ask for the working before the conclusion. Quotes before themes, definitions before numbers.

These prompts work in Claude and ChatGPT. They are illustrative starting points: change the brackets, keep the evidence rules.

Half of what goes wrong in a PM prompt is missing context: which customers you serve, what "active" means here, what the strategy says. Pasting that into every chat is tedious, so people stop doing it.

Put it in a Claude project or a ChatGPT project instead. Both keep files and custom instructions alongside the chats. Claude projects are on every plan (Free is capped at five), and ChatGPT projects take 25 files on Plus (Claude and OpenAI help centres). What to put in:

1. A metric glossary. One line each for activation, retention, active user, churned account, with the event and window. Link to your own metric definitions if you've written them down.
2. Personas or segments, with the words customers use, not the words marketing uses.
3. The current strategy or OKRs, so prioritisation prompts have something to score against.
4. Standing instructions, something like this:

You support the product team for [product]. Use the glossary in this
project for every metric. Never estimate a number: calculate it in code
from data I give you, or say you can't. When you summarise customer
material, attach participant IDs and verbatim quotes. When the material
doesn't answer something, write OPEN QUESTION.

Which assistant? For PMs we'd lean to Claude, because so much of the job is long reading, and Claude handles 20 files per chat and PDFs up to 1,000 pages. ChatGPT is a fine choice if your company already pays for it. The full comparison is in Claude vs ChatGPT.

Write an interview guide that doesn't lead

I'm interviewing [segment] about [problem area]. Draft a 30-minute guide.
Rules: no question may mention our product or a proposed solution; ask
about the last time they did [task], not what they usually do; include
two follow-ups per question that ask for a specific example.
Then list any of your questions that a participant could answer with
yes or no, and rewrite them.

The last line is the useful part: it makes the model audit its own leading questions.

Turn raw feedback into a countable list

Below are [n] support tickets and feedback comments, each with an ID.
Group them into problems. For each problem give the count of IDs, the
IDs themselves, and the two most specific quotes. Do not merge problems
unless the quotes describe the same failure. List IDs you couldn't place.

What to look for: an "unplaced" list. If every comment fitted neatly, the groups are probably too broad. Research on LLM coding found agreement with humans was high for broad themes and noticeably lower for specific ones (Liu and Sun, 2023).

Read open-text survey answers without inventing a percentage

Here are [n] open-text answers to "[question]". Classify each answer into
one of these codes: [codes], or "other". Return a table of answer ID and
code, then count the codes in code, not by estimating.

Asking for the row-level table first means you can spot-check ten rows. The percentage at the end is only as good as those rows.

Drafting a PRD from scratch is the obvious use. Critiquing one you've written is the better one, because you stay the author and the model does the nitpicking it's good at.

Critique a PRD

Review the attached PRD as a sceptical engineering lead. List: (1)
requirements that can't be tested as written, (2) assumptions stated as
facts, each quoted, (3) questions an engineer would ask in refinement.
Don't rewrite the document. Don't add requirements of your own.

Write acceptance criteria from a user story

User story: [story]. Context: [link or paste of the spec section].
Write acceptance criteria in Given/When/Then form. Cover the happy path,
empty states, permissions and errors. Mark any criterion that relies on
behaviour the spec doesn't describe with [ASSUMPTION].

Run a pre-mortem

It's six months after launch and [feature] has failed. Using only the
attached spec, research summary and metrics, list the five most likely
reasons. For each, name the evidence in the attachments that points to
it, or say "no evidence; speculation".

The label matters. A pre-mortem full of speculation is still useful, as long as it's marked as speculation and doesn't end up in a risk register as fact.

These are the prompts where a wrong answer looks most like a right one. A number comes back, nicely formatted, and nothing tells you it counted the wrong users.

Define before you calculate

I want [metric] for [period]. Before any calculation, write out:
the event or condition that counts, the time window, the denominator,
how duplicate or test accounts are excluded, and where your definition
differs from the glossary in this project. Wait for me to confirm.

Monthly active users is a good example of why. Google Analytics 4 counts MAU over a rolling 30 days, not a calendar month (GA4 Help), and duplicate accounts can inflate it: Meta once estimated about 11% of its worldwide MAUs were duplicates (Meta 10-K, FY2021).

Read out an A/B test

Here are the experiment results: [paste table]. Calculate the difference
and a confidence interval in code, and show the code. Then tell me what
the result does NOT show, including any segment where the sample is too
small to say anything.

Write funnel SQL against your schema

Schema: [paste table and column names]. Write SQL for a weekly funnel
from [event A] to [event B] within [n] days. State every assumption about
the columns. Don't use a column that isn't in the schema.

Then run it yourself. The best agent on the Spider 2.0 enterprise SQL benchmark solved 21.3% of tasks in the original paper (Lei et al., 2024), and real schemas are messier than any benchmark. More patterns in prompts for data analysis.

A good prompt changes what the model is willing to say. It doesn't make what it says true. Anthropic is direct about this: its techniques "don't eliminate" hallucinations, so "always validate critical information".

Three PM-specific failures survive good prompting:

• Quotes that are nearly right. A paraphrase in quotation marks reads as evidence. Search for it.
• Definitions that drift mid-conversation. The model agrees to your activation definition, then uses a looser one three answers later.
• Synthesis on top of sound parts. Every quote checks out; the conclusion strung between them doesn't follow.

Before you act on an answer:

1. Search for two quoted phrases in the source material.
2. Ask: "Which definition did you use for [metric], and where in this conversation did I confirm it?"
3. Ask: "Which of your conclusions rests on fewer than three pieces of evidence?"
4. Rerun the prompt in a new chat. Anthropic suggests comparing runs: "Inconsistencies across outputs could indicate hallucinations."

To do this claim by claim, connect Count to Claude or ChatGPT and ask it to audit the answer. Each claim is labelled as a citation, an assumption, model knowledge or a hallucination, with its reasoning chain. The wider setup is on our AI for product managers page.

How do product managers use ChatGPT?

Is Claude good for product managers?

Can AI write a PRD?

What's the best prompt for summarising user interviews?