Why your AI output is generic, and the PM fix
AI output feels generic because the model doesn't know your product, users, goals or constraints, so it fills the gaps with the most average answer. The fix is persistent context, not longer prompts: keep a few short files about your company, product, team, goals and constraints where your AI tool reads them automatically, make it ask questions before it drafts, and run the draft past a reviewer that attacks weak assumptions. Prompt tweaks help a little. Context and review change the output.
Why AI output is generic: three causes
1. No standing context
Every new chat starts from zero. Unless you paste it in, the model doesn't know:
- who your users are and which segment you're fighting for this year
- what's on the roadmap and what was cut last month
- the constraints: the platform you can't touch, the compliance rule, the two engineers you actually have
- what you've already tried and ruled out
With none of that, "write a PRD for saved searches" gets the PRD for saved searches at an average SaaS company. It reads fine. Your eng lead can't build from it, and your VP can't decide from it.
2. It answers before it understands
A one-shot prompt gets a one-shot answer. A human PM handed the same request would ask four questions first: for whom, why now, what does success look like, what's out of scope. The model answers immediately unless you tell it not to.
3. Nobody pushes back
Models tend to agree with the framing they're given. Ask for a launch plan and you get a launch plan, not "your pricing assumption is untested." The first person to attack the draft is usually a stakeholder in a meeting.
Fix 1: give the AI your context once
This is what developers call context engineering. For a PM it means five short files, each under a page.
| File | What goes in it (3–5 bullets) |
|---|---|
company.md | What you sell, to whom, business model, stage, the one metric leadership watches |
product.md | Core user jobs, main surfaces, what's live vs in beta, known weak spots |
team.md | Who's on the team, who decides what, your eng and design capacity |
goals.md | This quarter's goals and how they're measured; what you're explicitly not doing |
constraints.md | Technical, legal, budget and timing limits; decisions already made |
Where the files live, by tool:
- Claude Code: a
CLAUDE.mdfile in the project folder loads at the start of every session and can point to your other files (Anthropic's memory docs). - Codex, Cursor and other agents:
AGENTS.md, an open format many agents read (agents.md). Cursor also has project rules. - Claude or ChatGPT Projects: upload the files to the project and reference them in the instructions.
The AGENTS.md guide has a PM template you can copy.
Keep it current. When a decision changes, change the file. If your notes say "Quick update: leadership killed mobile and moved us to retention", that's three edits: goals.md (retention, not mobile), constraints.md (no mobile work this half) and a line in your decision log with who decided and when. A goals file that still says "grow mobile" will produce drafts that argue for mobile.
Don't dump everything. More context isn't better. Chroma's Context Rot report tested 18 models in 2025 and found performance "grows increasingly unreliable as input length grows", even on simple tasks. Earlier research (Lost in the Middle) found models use information at the start and end of a long input better than information in the middle. Five tight files beat a 40-page strategy deck.
Fix 2: make it ask before it answers
Add one instruction to your context file or the top of your prompt:
Before drafting anything, ask me 3–5 questions about the parts you'd otherwise have to guess: target user, the problem's evidence, success metric, scope limits, and dependencies. Wait for my answers.
Illustrative exchange (fictional product):
You: Draft a PRD for saved searches in Fieldnote.
AI: Before I draft: (1) Which users ask for this, field researchers or the ops managers who review their notes? (2) What's the evidence: tickets, interviews, usage data? (3) What would success look like in 60 days? (4) Is mobile in scope? (5) Does search share a backend with the notes export that's being rebuilt?
Each answer is a fact the model would otherwise have invented. Two minutes of questions saves a rewrite.
Fix 3: add a reviewer that attacks the draft
Run the draft through one or more reviewer passes before a human sees it. Use a fresh chat or a sub-agent so the reviewer isn't invested in the draft.
Find the single weakest assumption in this PRD. Explain what happens to the plan if it's wrong, and name the cheapest test that would tell us.
Review this as the eng lead. List what's underspecified, the riskiest technical dependency, and anything that will take 3x longer than it looks.
Read this as the target user described in product.md. Would you notice this change? What would you still complain about after it ships?
Packaged options ship these reviewers already. phuryn's free pm-skills has /red-team-prd and /pre-mortem. Anthropic's Product Management plugin has /brainstorm for testing assumptions. My AI PM OS has 12 review agents, including a devil's advocate and a customer-voice reviewer, and its PRD workflow runs a strengths review, then a flaws review, then revises.
Before and after: the same PRD prompt
Illustrative example. Fieldnote is a fictional B2B note-taking app for field researchers. Both outputs were written for this page to show the pattern; they are not a benchmark.
Prompt: "Write the Target users and Problem sections of a PRD for saved searches."
Target users: Busy professionals who need quick access to information and want to save time searching. Problem: Users waste time repeating the same searches. Without saved searches, they struggle to find important content quickly, which reduces productivity and satisfaction.
Target users: Ops managers at research agencies (our paying admins, ~20% of seats) who review field notes weekly. Not field researchers, who search rarely and mostly on mobile, which is out of scope this half. Problem: Managers rebuild the same 3–4 filters every Monday (site + date range + tag "follow-up") to find notes needing review. Evidence: 14 support tickets since July and 4 of 6 manager interviews. Retention goal: managers are the role that renews; review friction shows up in churn interviews. Constraint: search shares a backend with the export rebuild, so v1 saves filters client-side only.
The second version names a segment, says who's excluded, gives evidence, ties to a goal and states a constraint. Every one of those came from a context file or an answer to a question, not from a better prompt.
What doesn't work (or barely helps)
- "Act as a senior PM." A role prompt changes tone. It doesn't tell the model what your product is.
- Longer prompts with no new facts. Adjectives like "detailed, specific, strategic" don't add information. A sentence about your users does.
- Humanizer tools. They rewrite word choice so text sounds less like AI. A generic PRD that sounds human is still generic. They fix voice, not substance.
- Prompt libraries on their own. A good template from free prompts helps structure the output. It still needs your context to fill it.
A 30-minute setup checklist
- Pick one tool you'll use for PM drafts: Claude Code, Codex, Cursor, or a Claude/ChatGPT Project. (Claude Code setup guide.)
- Create a folder for your product work.
- Write the five context files above, 3–5 bullets each. Stop at one page per file.
- Add a
CLAUDE.mdorAGENTS.mdthat says: read the context files first, ask 3–5 questions before drafting, and flag any assumption you had to make. - Save the three reviewer prompts as files you can reuse.
- Test on a real artifact. Re-run a PRD section you wrote last month and compare. The PRD guide has a full workflow.
- Set a reminder to update the files when a decision changes.
If you'd rather start from a pre-built context system, AI PM OS ships one (five files, set up with /start). My free sample skills are at github.com/gnurio/nurijanian-skills. To see where your current setup is weakest, try the AI PM Setup Scorecard.
FAQ
Why do AI-generated PRDs all sound the same?
The model doesn't know your users, goals or constraints, so it writes the most probable PRD for the request. Without your context, the most probable answer is the average one.
Is a better prompt enough to fix generic AI output?
Not on its own. Better prompts improve structure and tone. Specific output needs specific facts, which come from context files, answers to the model's questions, and a review pass.
What is context engineering for product managers?
Writing down the facts an AI needs about your company, product, team, goals and constraints, and keeping them where your tool loads them automatically, such as CLAUDE.md, AGENTS.md or a Project.
What should I put in a CLAUDE.md or AGENTS.md file as a PM?
Pointers to your five context files, plus working rules: read context first, ask questions before drafting, flag assumptions, and the output formats you use for PRDs and updates.
How much context is too much?
When it stops fitting on a page per topic. Research such as Chroma's 2025 Context Rot report found model reliability drops as input grows. Keep files short and current.
Should I use ChatGPT or Claude Projects, or a file-based agent like Claude Code?
Projects work for chat-based drafting. File-based agents read and write files in your folder, so drafts and decisions accumulate. Either works if the context files exist.
Is it safe to put confidential product context into AI tools?
Your context goes to the tool's model provider under that provider's terms. Follow your company's AI policy and leave out anything it prohibits.
How do I keep context files up to date?
Edit them when a decision changes, not on a schedule. Some setups capture decisions from meeting notes and ask before writing to your files.
How do I make AI push back on my thinking?
Ask it to find the weakest assumption, review as your eng lead, or read as your customer. Use a fresh chat or sub-agent so the reviewer isn't defending its own draft.
Do humanizer tools fix generic AI output?
No. They change word choice and rhythm. They don't add facts about your product, so the content stays generic.