You just finished your eighth user interview this sprint. You've got 12 hours of transcripts, a wall of sticky notes, and a PRD due Friday. The synthesis alone will eat two full days.

Or you could build a prompt chain that handles it in 20 minutes.

This isn't about replacing your product thinking. It's about offloading the mechanical parts (coding transcripts, extracting patterns, formatting stories, writing boilerplate sections) so you spend your time on the decisions that actually matter.

Why Single Prompts Fail for PM Work

Most "AI prompts for product managers" articles give you a one-shot prompt: "Write a PRD for a mobile checkout feature." You paste it in, get back something that looks like a PRD, and realize it knows nothing about your users, your constraints, or your tech debt.

The output is structurally correct and contextually useless.

The fix is chaining. Each prompt in the chain does one job, and its output feeds the next step. Research synthesis feeds theme extraction. Themes feed user stories. Stories feed acceptance criteria. Criteria feed the PRD. At every handoff, you review, edit, and inject context the model doesn't have.

This mirrors how experienced PMs actually work. You don't jump from raw interviews to a finished PRD. You process information in stages. The chain just makes each stage faster.

Step 1: Interview Transcript Analysis

The biggest risk here: LLMs fabricate quotes. Multiple researchers have documented this. You ask for themes supported by quotes, and the model returns plausible-sounding phrases that don't exist in the original transcripts. Always verify quotes against your source material.

Here's a prompt that reduces hallucination by forcing the model to reference specific transcript locations:

You are a UX researcher analyzing interview transcripts. I will provide
transcripts from 6 user interviews about [onboarding experience for
our B2B analytics platform].

For each transcript, extract:
1. Direct quotes (verbatim, with speaker ID and approximate timestamp)
2. Pain points mentioned (quote the exact words used)
3. Feature requests or workarounds described
4. Emotional indicators (frustration, confusion, delight)

Rules:
- Only use exact text from the transcripts. Do not paraphrase or
  synthesize quotes.
- If a point is implied but not explicitly stated, flag it as
  "inferred" and explain your reasoning.
- Tag each finding with the participant ID.

[Paste transcripts here]

This works because it demands verbatim quotes with source attribution. The "inferred" rule gives the model an escape valve instead of forcing it to fabricate a supporting quote when one doesn't exist.

Expected output: A structured table per participant with 8-15 findings each, tagged with IDs like P3-14:32, containing exact transcript language. Inferred items clearly marked. No invented quotes.

You'll still need to spot-check 20-30% of the quotes against your transcripts. Trust but verify.

Step 2: Theme Extraction

Now you take that coded output and look for patterns across participants.

Below are coded findings from 6 user interviews about B2B analytics
onboarding. Each finding includes a participant ID, direct quote, and
category (pain point / feature request / workaround / emotional).

Group these findings into 3-7 themes. For each theme:
- Name it in 5 words or fewer
- List which participants mentioned it (by ID)
- Include 2-3 supporting quotes (verbatim from the findings above)
- Rate confidence: HIGH (4+ participants), MEDIUM (2-3), LOW (1)
- Note any contradictions between participants on this theme

Do not invent themes supported by only inferred findings.

[Paste Step 1 output here]

The participant count requirement matters. A theme mentioned by one person is an anecdote. A theme mentioned by five is a signal. Forcing the model to show its math lets you catch weak patterns before they become product decisions.

Expected output: 4-6 themes with names like "Dashboard setup requires external help" or "Data import expectations mismatched," each backed by specific quotes from 2-5 participants, with contradictions noted where they exist.

Step 3: User Story Generation

This is where most PMs start their AI usage. Starting here without the research steps means your stories are based on assumptions instead of evidence.

You are a product manager writing user stories for [B2B analytics
platform onboarding]. Use the research themes below as your source
material.

For each theme, generate 2-4 user stories in this format:
As a [specific user type from the research], I want to [action],
so that [outcome tied to a real pain point from the interviews].

Rules:
- Every story must trace back to a specific theme and quote from
  the research.
- User types must match the actual personas in the transcripts,
  not generic labels.
- "So that" must reference a real outcome participants described,
  not an assumed business benefit.
- If a story requires assumptions beyond the research, flag it
  with [ASSUMPTION] and state what you're assuming.

[Paste Step 2 output here]

Expected output: 10-18 user stories grouped by theme. Each story references specific participants. Stories flagged [ASSUMPTION] where research doesn't fully support the "so that" clause. User types like "first-time admin setting up a team workspace" instead of generic "user."

Step 4: Acceptance Criteria in Gherkin

Gherkin format (Given/When/Then) does double duty: it removes ambiguity for your engineers AND enables automated test scaffolding later. A good rule of thumb: 3-6 scenarios per story. If you need 12+, you're holding an epic, not a story.

Write acceptance criteria in Gherkin format for each user story below.

For each story, provide:
- 3-6 Given/When/Then scenarios covering the happy path, one edge
  case, and one error state
- Each scenario must be testable by a QA engineer with no additional
  context
- Use concrete values (specific numbers, names, states) not
  placeholders

Stories:
[Paste Step 3 output here]

Expected output: Structured Gherkin blocks like:

Given an admin has uploaded a CSV with 500+ rows
When the import completes processing
Then the system displays a summary showing rows imported, rows skipped, and specific error descriptions for each skipped row

Step 5: PRD Assembly

Now you have research-backed stories with testable criteria. The PRD prompt pulls it all together, but you need to inject the context no model has: your company's constraints.

Draft a PRD for [Onboarding Redesign - B2B Analytics Platform] using
the user stories and acceptance criteria below.

Product context (do not deviate from these constraints):
- Series B company, 12 engineers, shipping quarterly
- Primary users: analytics team leads at mid-market companies
  (200-2000 employees)
- Current onboarding completion rate: 34%
- Target: 60% completion within 90 days of launch
- Technical constraint: must work within existing React/PostgreSQL
  stack, no new infrastructure
- This is a Phase 1 scope. Defer anything requiring ML or third-party
  integrations.

Include these sections: Problem Statement, Goals and Success Metrics,
User Stories (grouped by priority), Scope (in/out), Technical Notes,
Open Questions.

Do NOT include sections for timeline, team allocation, or budget.

[Paste Steps 3-4 output here]

The product context block is the most important part of this prompt. Without it, you get a generic PRD for a generic company. With it, the model can make reasonable scoping decisions and flag stories that violate your constraints.

Expected output: A 3-5 page PRD with your specific metrics, constraints, and stories organized by priority. An "Open Questions" section that surfaces genuine ambiguities from the research (contradictions between participants, assumptions that need validation). Scope section that clearly defers items outside Phase 1.

The 70/30 Split

Practitioners who've run this workflow consistently land on the same ratio: AI gives you about 70% of the structure, and you bring 30%. That 30% is the hard part. It's knowing which theme actually matters for this quarter. It's recognizing that the story your loudest customer wants conflicts with your platform direction. It's cutting scope when the PRD looks clean but the engineering reality says otherwise.

The prompt chain handles transcription coding, pattern formatting, story structure, and section boilerplate. You handle judgment, priority, and the decisions that make a product good instead of just documented.

Some veteran PMs argue you shouldn't offload any of the writing because the act of writing forces you to think. There's truth in that. The counter: you're still thinking at every handoff point in the chain. You're reviewing themes, rewriting stories, adjusting criteria, and rewriting PRD sections. The chain doesn't remove your judgment. It removes the mechanical formatting work that sits between your decisions.

Make It Yours

These five prompts are a starting framework. The real value comes when you customize them for your product context, your user segments, and your team's conventions. Save your product context block as a reusable snippet. Build a library of your Gherkin patterns. Adjust the story format to match what your engineers actually want to read.

The PM who copies these prompts verbatim gets a decent first draft. The PM who adapts them into a personal system gets hours back every week.

If your product team wants to build prompt workflows like this across research, planning, and execution, Kief Studio runs hands-on training sessions tailored to your stack and process. Join the conversation on Discord (https://discord.gg/JfjyUdjJgP) or book a session at kief.studio/contact.

user research synthesis

Prompts for Product Managers: From User Research Synthesis to PRD Generation