For product managers
AI for product managers, from discovery to PRD
ZeroTwo is a place to think with several models over your own research. Paste interview notes, keep the evidence next to the claims, and turn a decision into a PRD an engineer can test against. The model drafts and challenges. You decide.
Free plan: 100 free credits to start, no card required.
The trail this page walks through
Evidence
Interview notes, quoted and numbered
Interpretations
Two readings that disagree
Decision log
What you chose, and what you are not doing
PRD
Each requirement points back to evidence
Acceptance criteria
Given / When / Then, including failure paths
Open assumptions
What would change your mind
60+ models from providers such as these sit behind the same chat, so you can compare drafts across models. Which ones you can pick depends on your plan.
- OpenAI
- Anthropic
- xAI
- Mistral
- DeepSeek
- Perplexity
- Qwen
Where the model helps, and where it should not decide
Most PM advice about AI stops at “it writes faster”. The useful line is between drafting and deciding.
Hand to the model
- Cluster raw interview notes by theme and quote the line each theme came from
- Argue the opposite reading of the same evidence
- Draft a PRD skeleton from a decision you have already made
- Turn each requirement into Given/When/Then cases and edge cases
- Rewrite the same decision for engineering, design or an executive
Keep for yourself
- Which customer problem is worth solving this quarter
- Reach, impact and effort inputs: the model has not seen your analytics
- What you are willing not to build, and for which segment
- Sign-off on every acceptance criterion
- Anything a stakeholder will hold you to later
Our view: a PRD is a record of a decision, not a generated document. If you cannot point from each requirement back to a piece of evidence, the draft is not ready, however polished it reads.
One decision, traced from interview to acceptance criteria
Six steps with the prompt for each, who does the work, and what you still check.
Collect evidence
You supply itPrompt you would send
Here are my interview notes, numbered by interviewee. Group them into themes. For each theme, quote the exact line it came from and the interview number. Do not paraphrase and do not add anything that is not in the notes.
- Interview 2
- “I photograph receipts and then forget them in my camera roll.”
- Interview 4
- “I invoice in euros twice a year. The conversion is annoying but rare.”
- Interview 5
- “I stopped using the app in March because month-end felt like homework.”
You still check: Anonymise names before pasting. Check each quote against your original notes: the model is told to quote, but you verify.
Argue two readings
Model drafts, you judgePrompt you would send
Give me two readings of this evidence that disagree with each other. For each, say which interview lines support it and what extra evidence would settle it.
- Reading A
- Capture friction drives drop-off, so receipt capture should come first. Supported by interviews 2 and 5.
- Reading B
- “Homework” is about reconciliation, not capture, so a capture feature will not fix it. Supported by interview 5 only.
- Gap
- Three notes cannot separate A from B. A follow-up question to the same people can.
You still check: Disagreement between readings is a prompt to go back to the evidence, not a vote to count.
Write the decision log entry
You decidePrompt you would send
Format this decision as a log entry: decision, reason, evidence used, what we are not doing, what would make us revisit. Use only what I wrote below.
- Decision
- Run receipt capture as a time-boxed experiment. Defer multi-currency.
- Reason
- Two of three notes point at capture; the third is ambiguous. Multi-currency was described as rare.
- Evidence
- Interviews 2 and 5 (weak), 4 (for deferral).
- Revisit when
- The follow-up interviews separate reading A from reading B.
You still check: This step has no model draft on purpose. The log records your judgment; the model only tidies the wording afterwards.
Draft the PRD from the decision
Model drafts, you checkPrompt you would send
Write a PRD section from this decision log entry. Include problem, goal, non-goals and numbered requirements. After each requirement, add “Evidence:” with the interview it traces to. If a requirement has no evidence, mark it “assumption”.
- Problem
- Users lose receipts between purchase and month-end.
- Non-goals
- Multi-currency. Bank feeds. Automatic categorisation.
- R1
- A user can attach a receipt photo to a transaction from the mobile app. Evidence: interview 2.
You still check: The “assumption” marker is the useful part of the prompt: it stops the draft from sounding more certain than your evidence.
Write testable acceptance criteria
Model drafts, engineering reviewsPrompt you would send
For requirement R1, write acceptance criteria in Given/When/Then form. Include at least one failure path and one permission case. Flag any criterion that needs a number from engineering instead of inventing one.
- AC1
- Given a transaction with no receipt, when the user takes a photo from the transaction screen, then the image is attached and the transaction shows a receipt marker.
- AC2
- Given camera permission is denied, when the user taps Add receipt, then the app explains how to enable it and offers file upload instead.
- AC3
- Given an attachment that cannot be stored, when the user confirms, then an error is shown and the transaction is unchanged.
You still check: Engineering owns limits such as file size. Leave them as open questions until someone with the numbers answers.
List what is still unknown
You own itPrompt you would send
List every assumption in this PRD that is not backed by one of the numbered interviews, and for each, the cheapest way to check it.
- Segment
- Do these three people represent the segment we are targeting?
- Cause
- Is capture or reconciliation the real driver of drop-off?
- Effort
- How long does R1 take? Only engineering can estimate this.
You still check: A PRD that ships with an honest unknowns list is easier to defend than one that implies everything was settled.
Handing the finished PRD to engineering? The build side is covered on AI for developers.
Prioritisation: you supply the inputs, the sheet does the arithmetic
RICE is reach × impact × confidence ÷ effort. A model can format the table and question your numbers. It cannot know them.
| Backlog item | Reach (users per quarter) | Impact (0.25 to 3) | Confidence | Effort (person-months) | Score |
|---|---|---|---|---|---|
| Receipt capture | 400 | 1 | 50% | 2 | 100 |
| Reconciliation rewrite | 400 | 2 | 30% | 4 | 60 |
| Multi-currency | 60 | 2 | 80% | 3 | 32 |
All inputs are invented for the imaginary app above. They are not data from a real product.
What the table shows
The reconciliation rewrite has the same reach as capture and twice the impact, yet it ranks lower because confidence is 30%. The honest next step is more interviews, not a bigger estimate.
What to ask the model
“Which of these inputs rests on the weakest evidence, and what would raise confidence?” Use the answer to plan research, then enter your own numbers.
Which ZeroTwo tool for which PM job
Each row names what the tool does for you and the part you still check.
| PM job | Feature | What it does for you | You still check |
|---|---|---|---|
| Synthesise interviews or a research PDF | Chat with uploaded files | Ask for themes with the source line and interview number for each. | Every quote against your original notes. |
| Draft and revise the PRD | Canvas | Keep the PRD as a document you edit next to the conversation. | Non-goals and open questions. These stay yours. |
| Scan the market for an input to a decision | Web search and deep research | Gather sources you can open. For SaaS markets, see the worked method on SaaS competitive research. | Open each cited source. Competitor pricing and features change. |
| Pressure-test a reading | Switching models | Run the same prompt on two models and compare where they disagree. | Disagreement points you back to the evidence; it is not a vote. |
| Stakeholder update | Chat | Rewrite the decision log for engineering, design or an executive. | Dates, numbers and commitments before you send. |
What a PM needs from a plan
You can try the whole worked example on Free. Paid plans add credits and tools; prices are from the pricing page.
Free
$0/mo
Enough to run the worked example on one real decision.
- 100 free credits to start
- +20 bonus credits every day you log in
- 1 file upload per message
- Serper web search
Plus
$14.99/mo
More credits, multiple file uploads per task, and the Work agent.
- 1,650 monthly credits
- +20 bonus credits every day you log in
- Multiple file uploads per task
- Access to Work and Code agents
- Access to deep research
Pro
$29.99/mo
$26.99/mo billed annually
Adds Bots, unlimited projects and a larger monthly credit pool.
- 3,500 monthly credits
- +30 bonus credits every day you log in
- Unlimited projects
- Access to Bots
Full comparison on the pricing page, or compare AI plan pricing. Weighing it against a single-model subscription? Read how ZeroTwo differs from ChatGPT.
Questions product managers ask
Can AI write a good PRD?
It can draft one quickly and it is useful for structure, but a PRD is only as good as the decision and evidence behind it. Give the model the decision, the numbered evidence and the non-goals, then check every requirement against a source. Without those inputs you get a document that reads well and that nobody can defend.
Which model should I use for product work?
ZeroTwo does not publish a ranking, because the right model depends on your material and your taste. Run the same prompt on two models and compare the drafts and where they disagree. Which models you can pick depends on your plan: Free has limited model selection.
Can AI do the prioritisation for me?
It can do the arithmetic, format a RICE or MoSCoW table, and challenge your inputs. It cannot supply reach, impact or effort, because it has not seen your analytics or your engineers' estimates. Confidence is the input most worth arguing about, and it is yours.
How do I avoid invented customer quotes?
Paste real notes and ask for themes with the exact quote and interview number. Then check each quote against the original. Never ask a model to generate customer feedback and file it as research.
Is it safe to paste customer interview notes?
Check your company's policy on sharing customer data with any AI service, and read the ZeroTwo privacy policy. Anonymise names and identifying details before pasting, as the worked example above does.
What does it cost?
There is a Free plan with 100 free credits to start, 20 bonus credits every day you log in, and no card required. Plus is $14.99 per month and Pro is $29.99 per month ($26.99 per month billed annually). Deep research is included from Plus upward (Plus: 150 a month, using credits).
Go deeper
- PRD templateA starting structure for the document in step 4.
- Sprint planning with AICarry accepted requirements into a sprint.
- Customer interviews into a PRDA longer write-up of the evidence-to-requirements workflow.
- AI for developersWhere implementation starts, once the PRD is handed off.
Also useful around the edges of the job: if you are hiring a PM, see the recruiting email templates; if you market a SaaS product, the SaaS content marketing guide. More tools are on AI tools and the AI text generator.