A prompt template for product managers: role, context, objective, format

Why does a product manager need a prompt template?

A one-line prompt makes the model guess the audience, the purpose, the length and the standard, and it guesses in the direction of generic. Cigdem Cevrim, a senior product manager at STRV, puts the problem in two sentences: "Give an AI model structure and it performs. Give it an unclear request and it guesses confidently" (STRV).

Product work makes the guessing worse. A PRD section, a research synthesis and a release note all have readers with different needs, house formats and a cost to being wrong. A template does not make the model smarter. It makes you write down the decisions you would otherwise leave to it.

What does the template look like?

Six labelled sections, in this order. Copy it, keep the labels, and replace the bracketed text.

ROLE
You are [who the model should act as], writing for [the reader].

CONTEXT
[What the model cannot know: the product, the users, the situation,
the constraints. Paste source material here, inside delimiters.]
"""
[pasted notes, data, transcript, draft]
"""

OBJECTIVE
[One sentence: what the output is for and what decision it supports.]

FORMAT
[Structure, headings, length, tone. What to leave out.]

CHECKS
- Use only the material in CONTEXT. If something is missing, say so.
- [Any rule the output must pass before you would send it.]
- Before answering, list anything in the task that is unclear.

EXAMPLES
"""
[One or more real examples of the output you want.]
"""

Each section exists because a model answers a specific question badly when it is missing.

Section	Question it answers	What goes wrong without it
Role	Whose judgement and vocabulary to use	Answers pitched at nobody in particular
Context	What is true about this product and situation	Plausible filler instead of your facts
Objective	What the output is for	A summary when you needed a recommendation
Format	What shape to deliver	Walls of prose, or bullets where you needed a table
Checks	What the output must pass	Invented numbers, skipped gaps, unflagged assumptions
Examples	What good looks like here	Correct content in a format your team never uses

Context does most of the work. Anthropic's documentation recommends 3 to 5 examples when you use them (Anthropic prompting best practices), but a prompt with thin context and five examples still produces confident nonsense about your product.

Where does this template come from?

The six-section template is an extension of RODES, a public framework whose letters stand for Role, Objective, Details, Examples and Sense check. STRV's version describes the last step as "Ask if the task is clear" (STRV), and the juuzt.ai knowledge base lists the same five components with "Sense Check" as a final review of the response.

The template here makes two changes for product work. It splits Details into Context and Format, because those are the two things PM prompts leave out most, and they fail differently. It turns the sense check into a Checks section with rules the output must pass, and keeps "list anything unclear" as one of them, so the model asks before it guesses.

How does the template change a PRD section prompt?

Before:

Write the problem statement for our new bulk-edit feature.

After:

ROLE
You are a senior product manager writing for engineers and a design lead.

CONTEXT
We sell scheduling software to dental clinics. Clinic managers edit
appointments one at a time. Support notes and two interview summaries:
"""
[pasted notes]
"""

OBJECTIVE
A problem statement that lets engineering judge scope before we
commit to a solution.

FORMAT
Under 150 words. Who has the problem, when it happens, what it costs
them, and how we know. No solution ideas.

CHECKS
- Every claim must trace to the pasted notes. Mark anything that does not.
- If the notes do not say how often this happens, write "frequency unknown".

What changed is not the length. The second prompt tells the model who reads the output, forbids solutions (the thing a one-line prompt always adds), and makes missing evidence visible instead of papered over. The PRD template shows where that problem statement sits in the full document.

How does it work for interview synthesis?

Before:

Summarise these customer interviews.

After, showing only the sections that changed:

CONTEXT
Six 30-minute interviews with finance leads at mid-size retailers,
about month-end reconciliation. Transcripts:
"""
[transcripts]
"""

OBJECTIVE
Find the problems worth a discovery sprint, not a summary of each call.

FORMAT
A table: problem, how many of the six interviews mentioned it,
one verbatim quote, and whether they described a workaround.

CHECKS
- A problem mentioned by one person is listed separately as "single mention".
- Quote exactly. Never merge two people's words into one quote.

The count column and the "single mention" rule stop the most common synthesis failure, where one vivid complaint reads like a pattern. For the method underneath, see research synthesis.

How does it work for release notes?

Before:

Write release notes for these tickets.

After:

ROLE
You write release notes for customers who use the product daily and
do not read engineering tickets.

CONTEXT
"""
[list of merged tickets with titles and descriptions]
"""

OBJECTIVE
Tell customers what changed for them this week.

FORMAT
Group by what the customer can now do. One sentence per change.
Skip internal refactors and anything with no visible effect.

EXAMPLES
"""
[last month's release note, the one your team was happy with]
"""

Here the Examples section earns its place, because release notes have a house voice that is faster to show than to describe. AI release notes for product managers covers the full workflow.

How should you format the sections so the model reads them correctly?

Separate instructions from material. Capitalised labels, triple quotes around pasted text, or XML-style tags all do the job. Anthropic's documentation explains why the separation matters: "XML tags help Claude parse complex prompts unambiguously, especially when your prompt mixes instructions, context, examples, and variable inputs" (Anthropic prompting best practices).

The failure this prevents is concrete. If you paste a customer email that says "please summarise this for my manager" without delimiters, the model may treat the customer's sentence as your instruction.

What should you do when the first output is mediocre?

Expect it, and fix the prompt rather than the output. If you edit the answer by hand, the next run makes the same mistake. If you add the missing rule to Checks or the missing fact to Context, every future run improves.

Two moves speed this up. Ask the model to critique your prompt before running it: which sections are vague, and what it would have to assume. STRV's guide recommends the same idea as meta-prompting, using a model "to improve the prompt itself." And keep the "list anything unclear" check, since a question from the model costs you ten seconds and a wrong assumption costs a rewrite.

When is a template overkill?

For a quick rewrite of one sentence or a factual lookup, six sections is ceremony. The template pays off when the output will be read by someone else, reused, or acted on. A reasonable rule: if you would review a colleague's draft of this before it went out, write the prompt with the template.

The other risk is over-constraining. A Format section that dictates every heading leaves no room for the model to notice that your data suggests a different structure. When exploring, loosen Format and tighten Checks.

Where does the AI Prompting for Product bootcamp fit?

AI Prompting for Product is a 1 week Builders Camp bootcamp with 2 live sessions and 19 self-paced microlessons, directed by Andre Albuquerque. Its public syllabus lists prompt structure (roles, context, constraints and success criteria), research and synthesis prompts, and writing and stakeholder communication prompts among its topics, and names a reusable prompt library as one of the things you gain.

See the AI Prompting for Product bootcamp

For the prompts behind one specific artifact, see AI prompts for writing user stories. For how a template becomes the standing instruction inside an AI feature, see system prompt, and for the discipline as a whole, prompt engineering. When the output format keeps drifting, few-shot prompting is the next tool.

Save each filled template next to the output it produced and the edit you made to it. After five or six uses, the edits tell you which rule belongs permanently in Checks, and the template becomes your team's standard instead of your own habit.