ChatGPT is not magic and not useless. The PMs who get real value from it treat it as a first-draft engine and thinking partner, not an oracle. The ones who waste time with it either expect too much or prompt too vaguely. The difference is almost always in how you use it, not which version you're running.
This guide is practical. Six real use cases, real prompts, and honest coverage of where the model falls down.
The 6 PM Use Cases That Actually Work
1. User Research Synthesis
Synthesizing interview notes is tedious and pattern-matching is where LLMs genuinely shine. Paste raw notes in, ask for themes and contradictions.
Prompt:
"I interviewed 12 users about [problem area]. Here are the raw notes: [paste]. Identify the top 3 themes across all interviews. Flag any contradictions or outliers. Do not summarize. Show me the patterns."
The output won't replace your judgment on which themes matter, but it will surface connections you might miss when you're deep in the weeds. Pair this with your own discovery process. AI synthesis is a starting point, not a conclusion.
2. PRD Drafting
ChatGPT can produce a credible one-page PRD skeleton faster than you can open a doc. The value is speed and structure, not correctness.
Prompt:
"Write a one-page PRD for [feature name]. Target users: [describe]. Problem being solved: [describe]. Proposed solution: [describe]. Success metric: [describe]. Format it with sections for problem, solution, success metrics, open questions, and out of scope."
You will need to add internal context, technical constraints, and stakeholder requirements the model cannot know. Use tools like Forge if you want AI document drafting with more product-specific structure built in.
3. Competitive Analysis First Pass
ChatGPT won't give you accurate pricing or live feature data, but it's useful for understanding public positioning and summarizing what a competitor claims about themselves.
Prompt:
"Based on publicly available information, summarize the key features and positioning of [competitor]. What problem do they claim to solve? Who is their stated target customer? What is their primary differentiator?"
Use this to orient yourself before doing real research. Do not use it as a source of record. Follow up by reading the actual website, G2 reviews, and any public pricing pages yourself.
4. Stakeholder Communication
PMs write a lot of updates that need to land differently depending on the audience. ChatGPT is good at register-shifting: taking a technical update and translating it for an executive.
Prompt:
"Rewrite this update for a CEO audience. Remove technical jargon. Lead with business impact. Keep it under 100 words. Here's the original: [paste]."
You can also ask it to harden weak language, remove qualifiers, or add specificity. This is one of the higher-ROI use cases because the raw material is yours and the model is just reshaping it.
5. Brainstorming Edge Cases
Before shipping, experienced PMs think about failure modes. ChatGPT is a fast way to stress-test a feature design.
Prompt:
"I'm building [feature]. What are 10 ways a user could misuse it, abuse it, or encounter unexpected behavior? Include both intentional misuse and unintentional user errors."
This pairs well with red teaming. If you're building AI-powered features, see the red teaming guide for a more structured approach to adversarial testing.
6. OKR and Success Metric Drafting
Metric-setting conversations often stall because no one wants to commit to a number. ChatGPT can propose measurable Key Results that spark the real discussion.
Prompt:
"Help me write 3 measurable Key Results for this Objective: [paste]. Make them outcome-focused, not output-focused. Each KR should have a clear numeric target and timeframe."
The model will produce something plausible. Your job is to anchor the numbers to actual baselines and push back on anything that looks like activity metrics dressed up as outcomes. Use the product metrics guide to sanity-check what you're measuring.
What ChatGPT Is Bad At
Be direct with yourself about the limits:
Real-time and recent data. The model has a training cutoff. Anything about current market conditions, recent product launches, or live pricing will be stale or fabricated.
Internal context. ChatGPT knows nothing about your users, your codebase, your team dynamics, or your company's strategic priorities. Every output that requires that context needs you to supply it, or the output is generic at best.
Accurate competitive specifics. It will confidently describe a competitor's pricing tier that doesn't exist, or a feature that was deprecated two years ago. Always verify.
Nuanced user empathy. The model has read a lot about users but has not talked to yours. It can pattern-match on common user psychology but will miss the specific friction points that your research surfaces.
Prompt Patterns That Work
Role assignment: Start prompts with a role. "Act as a senior PM at a B2B SaaS company" gives the model a persona that shapes tone and assumptions. More specific roles produce more useful outputs.
Context before task: Give the model what it needs to know before you tell it what to do. Reversed prompts ("Write a PRD for X, context: Y") produce worse outputs than front-loaded context.
Format specification: Ask for the specific format you want. "Bullet points only," "table with three columns," "no more than 200 words." Without format constraints, the model defaults to verbose prose.
Ask for alternatives: When a first output feels flat, ask: "Give me three alternative versions of that. One that's more direct, one that's more data-driven, and one that's more narrative." Alternatives reveal the prompt's range and give you better raw material to work from.
For deeper work on writing effective AI prompts, the prompt engineering guide for PMs covers structured prompting frameworks you can apply to both ChatGPT and any LLM your product uses.
The "Would I Stake My Career on This?" Filter
This is the most useful heuristic for evaluating AI output. Before you send a ChatGPT-drafted email, present an AI-synthesized insight, or include a model-generated metric in a roadmap review: ask whether you would stand behind it if questioned.
If the answer is "probably," that's not good enough. Edit until you can say "yes." That means:
- You've verified the facts independently
- You've added the internal context the model lacked
- You've removed any vague language that the model introduced
- The voice sounds like you, not like a confident AI approximation of you
ChatGPT is most useful when you treat it as a capable first-draft collaborator with significant blind spots. The better you get at compensating for those blind spots with your own judgment, the more value you extract.
Getting Started
If you haven't built a regular AI workflow yet: pick one use case from this list and commit to using it for the next five working days. User research synthesis is a good entry point because the output is easy to evaluate against your actual notes.
The goal is not to delegate your thinking. It's to spend less time on the parts of PM work that are mechanical and more time on the parts that require your specific knowledge of your users, your market, and your team.