Why this is hard to get right
Maya is a senior product manager at a mid-size SaaS company. She's just wrapped up two weeks of stakeholder interviews — 14 conversations with engineers, customer success leads, sales reps, and two VPs. Her notes fill a Notion doc that's nearly 4,000 words. The roadmap planning session is in 48 hours.
She opens a blank document and stares at the wall. She knows roughly what people said. She remembers the disagreements. But she can't see the shape of it. And she needs a one-pager by tomorrow morning.
Her first attempt is what most people try: she pastes her notes into ChatGPT and types, "Summarize my stakeholder interviews and tell me what the key takeaways are." The AI returns a clean-looking response — six bullet points, some language about "alignment" and "customer experience." It's coherent. It's also useless. None of the bullets are specific enough to drive a decision. There's no attribution, no tension, no prioritization. It reads like a consultant paraphrasing a memo they didn't write.
Maya tries again. She adds "be more specific" to her next message. The AI gets wordier but not sharper. She tries "focus on disagreements." It surfaces two vague conflicts. She pastes the notes again with "what should we do next?" The suggestions are generic — "invest in customer onboarding," "improve internal communication."
She's wasted 90 minutes and has nothing she can share with a VP.
The core problem isn't the AI. It's the prompt. The AI doesn't know who will read this. It doesn't know whether to surface three themes or ten. It doesn't know whether a disagreement between engineering and CS matters more than one between two sales reps. It doesn't know that Maya needs direct quotes to build credibility with skeptical stakeholders. And it certainly doesn't know that she needs effort and impact estimates tied to owner types so her roadmap session doesn't devolve into scope theater.
When Maya structures her prompt — defining her role, her audience, the exact output format, the quote rules, and the accuracy constraints — the AI produces something she can actually use. The first pass gives her 6 themes with quotes and attribution, 3 flagged disagreements with root cause hypotheses, and 5 next actions with PM/Eng/CS owners and effort tags. She edits for 20 minutes and sends it to her VP before dinner.
The difference between the vague prompt and the structured one isn't AI capability. It's information. A well-built prompt tells the AI exactly what "good" looks like — and that's the only way to get output that earns trust in a room full of decision-makers.
Common mistakes to avoid
Omitting Who Will Read the Output
When you don't specify your audience — VP Product vs. a cross-functional team vs. a board — the AI defaults to a generic professional register. The depth, vocabulary, and level of evidence change dramatically depending on who's in the room. Always name job titles and decision context. An exec readout needs conclusions up front; an engineering sync needs root causes and tradeoffs.
Asking for Themes Without Specifying How Many
Saying 'find the main themes' gives the AI license to return 3 or 15, with no calibration to your actual use. Too few themes collapse nuance; too many are unusable in a meeting. Specify a range like 5–7 and add a format requirement — what each theme must include — so the output is structurally consistent and scannable, not just a list of topic labels.
Skipping Attribution and Evidence Rules
A synthesis without quotes or source attribution is just an opinion. If the AI doesn't know you need direct quotes tied to specific stakeholder roles, it will paraphrase everything. Paraphrasing erases credibility. Tell the AI the quote length range, that quotes must be attributed to a role (not a name), and that it should flag low-confidence inferences rather than state them as facts.
Forgetting to Surface Disagreements Explicitly
Most prompts ask for consensus themes and miss the conflicts entirely. Unresolved disagreements are often the most decision-relevant content in an interview set. If you don't explicitly ask the AI to identify conflicts, name the parties, and hypothesize root causes, it will smooth over tensions to produce a tidy narrative — which is the opposite of what drives alignment.
Leaving Out Scope and Timeframe Filters
Interview notes accumulate context from many conversations, including legacy complaints, off-topic tangents, and historical gripes. Without a scope instruction — such as 'focus on the next 2 quarters' or 'exclude comments about the legacy billing system' — the AI may weight outdated issues equally with current priorities, muddying the output and misleading decision-makers.
Not Specifying Output Format or Length
Without format constraints, the AI chooses its own structure — and it rarely matches what you need. You might get a narrative essay when you need a table, or a bulleted list when you need owner assignments. Specify the exact sections, approximate length, and any visual structure like tables or priority tags. The more precise the format, the more directly the output plugs into your deliverable.
The transformation
Summarize my stakeholder interviews and tell me what the key takeaways are.
You’re a **product research lead**. Synthesize the stakeholder interview notes I paste below. 1) Create a **1-page readout** for **VP Product and Engineering managers**. 2) Identify **5–7 themes**, each with: what it is, **who said it**, and **1 direct quote**. 3) Call out **top 3 disagreements** and explain likely causes. 4) Recommend **5 next actions** with owner type (PM/Eng/CS), effort (S/M/L), and impact (H/M/L). **Tone:** neutral, decision-focused. **Constraints:** don’t invent details; flag missing info as questions. **Notes:** [paste notes]
Why this works
Role Anchors Judgment
The After Prompt opens with 'You're a product research lead.' This isn't decoration. It tells the AI what expertise to apply when making judgment calls — which conflicts matter, how to frame next steps, what a VP considers actionable. Without a role, the AI defaults to a generalist register that lacks the decisional weight the situation requires.
Structured Output Prevents Drift
The prompt specifies four distinct numbered sections: a one-page readout, themes with three required sub-elements, top disagreements with causes, and next actions with three tagging dimensions. Each section locks the AI into a deliverable shape. The AI can't collapse themes into vague bullet points or skip the disagreements section — the structure enforces completeness.
Evidence Requirements Build Credibility
The After Prompt requires each theme to include 'who said it' and '1 direct quote.' This single instruction transforms the output from AI-generated opinion into stakeholder-backed evidence. Quotes create trust in a readout. Decision-makers can evaluate the source, not just the conclusion. Without this rule, the AI summarizes and strips all traceability.
Action Tags Create Immediate Utility
The five recommended next actions must each carry an owner type (PM/Eng/CS), an effort rating (S/M/L), and an impact rating (H/M/L). This three-axis tagging system turns insights into a prioritization framework. The AI isn't just reporting what it heard — it's producing a decision-ready artifact that walks directly into a roadmap session.
Accuracy Constraints Prevent Hallucination
The final instruction — 'don't invent details; flag missing info as questions' — directly addresses the most dangerous failure mode in synthesis prompts. Without this guardrail, AI systems confabulate plausibly-sounding details that erode trust when they surface in meetings. Flagging gaps as questions is more useful than filling them with fabricated continuity.
The framework behind the prompt
The Research Behind Synthesis Prompts
Stakeholder synthesis sits at the intersection of qualitative research methodology and organizational decision-making. Understanding the theory behind it helps you build prompts that produce genuinely useful output — not just tidy summaries.
Thematic analysis is the foundational method here, developed by researchers Braun and Clarke in 2006. It involves reading across data sources to identify recurring patterns that represent shared meaning, not just frequency. A good theme is interpretive — it explains why something matters, not just that it appeared. Prompts that ask for topic labels rather than interpretive themes produce weaker outputs because the AI mimics the surface task without the analytical depth.
Affinity mapping, popularized in UX research and design thinking, organizes qualitative data into clusters before synthesis. When you pre-segment your notes by stakeholder group before running the AI prompt, you're applying the same logic — reducing cognitive load on the model and preserving signal from each source before aggregating.
The STAR framework (Situation, Task, Action, Result) appears implicitly in strong synthesis prompts. Next-action sections that require owner type, effort, and impact are essentially STAR structured for organizational decisions — they connect an insight to a situation, assign a task, and anticipate a result. This structure is why the After Prompt's next-action format produces planning-ready output rather than vague recommendations.
Organizational conflict theory explains why surfacing disagreements is as important as surfacing consensus. Research by Jehn (1995) distinguishes between task conflict (disagreements about work content) and relationship conflict (interpersonal friction). Task conflict, when surfaced and addressed, improves decision quality. Prompts that flatten conflict into consensus themes hide the most decision-relevant information in an interview set.
Finally, few-shot prompting — providing example outputs inside the prompt — dramatically improves synthesis quality for complex, format-sensitive tasks. If you've run this synthesis before, pasting a redacted example of a strong past output teaches the AI what "good" looks like far more precisely than any instruction alone.
Prompt variations
You're a customer success analyst. Synthesize the customer and internal interview notes I paste below into a QBR-ready briefing.
- Identify 4–6 themes from customer feedback. For each: state the theme, the customer segment affected, and one verbatim quote under 20 words.
- List top 3 risks the customer has raised, ranked by likelihood of impacting renewal.
- List top 3 opportunities for expansion or upsell, with the evidence behind each.
- Recommend 4 next actions for the CS team, each with a 30-day or 90-day timeframe.
Tone: customer-focused, honest about gaps. Constraint: do not soften negative feedback — flag it clearly.
Notes: [paste interview notes here]
You're a sales enablement strategist. Analyze the win/loss interview notes below from a recently closed deal.
- Identify the top 3 reasons we won or lost, each supported by a direct quote from the notes.
- List 5 objections raised during the deal with the stakeholder role that raised them.
- Identify 2–3 competitive factors that influenced the decision, based only on what was said.
- Recommend 4 enablement actions — talking points, collateral, or process changes — with the team responsible (AE, SE, Marketing).
Tone: direct, evidence-based. Constraint: distinguish between what the buyer said and what the seller inferred. Flag inferences clearly.
Notes: [paste debrief interview notes here]
You're a staff engineer facilitating a decision review. Synthesize the technical stakeholder interview notes below for an engineering leadership discussion.
- Identify 4–5 technical themes, each with the team or role that raised it and a representative quote.
- Surface top 3 tradeoffs where engineering, product, and operations expressed conflicting priorities. For each, state the positions clearly and the likely root cause.
- Flag any scope or dependency risks mentioned across multiple interviews.
- Suggest 3 clarifying questions the engineering leadership team should resolve before committing to a direction.
Tone: technically precise, neutral on outcomes. Constraint: do not recommend a solution — surface the decision space only.
Notes: [paste engineering interview notes here]
You're a chief of staff preparing a board-level briefing. Synthesize the strategic stakeholder interview notes below into a concise executive input document.
- Write a 3-sentence executive summary covering the top finding, top risk, and recommended focus area.
- Identify 3–4 strategic themes, each stated as a conclusion — not a topic — with one supporting quote.
- Highlight the single sharpest disagreement across stakeholders and explain what resolving it would unlock.
- Propose 3 strategic questions for the leadership team to discuss, ranked by urgency.
Tone: crisp, confident, board-appropriate. Constraint: no bullet soup — use full sentences for every finding. Flag anything that needs verification before the session.
Notes: [paste strategic interview notes here]
When to use this prompt
Product Managers running discovery
Turn stakeholder interviews into themes and an action plan before roadmap planning.
Customer Success leaders prepping QBRs
Synthesize internal and customer interviews into a clear set of priorities and risks.
Sales teams capturing deal feedback
Convert post-mortem interviews into objections, proof points, and enablement actions.
Engineering managers aligning on tradeoffs
Highlight disagreements and root causes so teams can resolve scope and priorities faster.
Executives reviewing strategic input
Get a one-page readout that surfaces alignment gaps and next steps without extra meetings.
Pro tips
- 1
Specify the decision this synthesis must support so the AI ranks themes by usefulness.
- 2
Define your stakeholder groups and titles so attribution and conflicts stay accurate.
- 3
Set a quote rule, like 8–20 words per quote, so the output stays scannable.
- 4
Add a timeframe and project scope so the AI filters out legacy issues and side topics.
When your interview set exceeds 15 conversations or 8,000 words of notes, a single-pass synthesis prompt loses precision. The AI starts averaging across too much input and surfaces the loudest signals rather than the most important ones.
Use a two-pass approach:
-
Pass 1 — Segment synthesis. Break your notes into groups — by stakeholder type, department, or customer segment. Run the core synthesis prompt on each group separately. Ask the AI to produce 3–4 themes per group with supporting quotes.
-
Pass 2 — Cross-segment synthesis. Paste the outputs from Pass 1 into a new session. Prompt the AI: 'You're a product research lead. Compare these segment-level syntheses. Identify 3 themes that appear across all segments, 2 themes unique to one segment (and explain why they matter), and the sharpest cross-segment disagreement with root cause hypotheses.'
This layered approach preserves segment-level nuance while producing a clean top-line synthesis. It's especially effective when you need to show executives both the overall pattern and the variation beneath it.
Tip: Save each segment synthesis as a named block before the cross-segment pass. Label them clearly — for example, 'Engineering synthesis,' 'CS synthesis' — so attribution survives across sessions.
The core synthesis structure — themes, disagreements, next actions — applies broadly, but the weighting changes by industry and research goal.
UX research teams should add a section for behavioral observations distinct from stated preferences. Stakeholders often say one thing and do another. Prompt the AI: 'Separate what stakeholders said they want from what their described behavior suggests they actually do. Flag any gaps between stated preference and implied behavior.'
Healthcare and regulated industries require a stricter accuracy constraint. Add: 'Do not infer clinical conclusions from qualitative interview data. Label all findings as preliminary observations requiring validation.' This protects against using synthesis outputs as clinical or compliance evidence.
Agency and consulting contexts often involve synthesizing across client stakeholders with competing agendas. Add: 'Where stakeholders represent different organizations or contracting parties, flag organizational affiliation alongside role. Distinguish between client priorities and vendor priorities in the disagreements section.'
Early-stage startups running founder-led discovery should add: 'Weight themes by frequency and by the stakeholder's proximity to the purchase or usage decision. A comment from a daily user outweighs a comment from a sponsor in most cases — flag when this weighting affects your findings.'
The quality of your synthesis output depends heavily on what you paste in. Use this checklist before running the prompt.
Before you paste:
- Replace all personal names with role labels (VP Engineering, Enterprise Customer, SDR)
- Add a source header before each interview block: role, date, interview type (in-person, remote, async)
- Remove embedded images, links, and table formatting that may confuse the model
- Delete internal commentary you added during review — the AI may treat your annotations as stakeholder input
- Check that each interview block is clearly separated, ideally with a line break and header
Scope filters to add to your prompt:
- Time horizon (e.g., 'focus on the next two quarters only')
- Project scope (e.g., 'limit synthesis to comments about the reporting feature')
- Exclusion rules (e.g., 'ignore comments about the legacy API unless they appear in 3 or more interviews')
After you get output:
- Verify every direct quote against your source notes before sharing
- Check that owner assignments make sense given your actual team structure
- Flag any theme that appears only once in the notes — the AI may have over-weighted a single strong voice
- Add a 'confidence level' annotation to any finding that rests on fewer than 3 interview sources
When not to use this prompt
This prompt pattern is not appropriate in every synthesis situation. Understand the boundaries before you build your workflow around it.
- When you have fewer than 4 interviews: The AI will generate themes from too thin a data set, overfitting to one or two strong voices. With small N, use the AI to organize notes — not to synthesize patterns.
- When your notes contain sensitive HR or legal content: AI synthesis of notes involving disciplinary issues, legal disputes, or protected-class discussions creates risk. These situations require human-only analysis and legal review.
- When the decision is already made: If leadership has committed to a direction and you're documenting post-hoc, a synthesis framed around "what should we do next" will produce misleading output. Use a documentation-oriented prompt instead.
- When you need statistical validity: Qualitative synthesis identifies themes, not statistical significance. If your stakeholders represent a sample you intend to generalize from, you need quantitative methods — not an AI thematic summary.
- When notes were taken inconsistently: If some interviews are verbatim transcripts and others are sparse bullet points, the AI will weight the detailed notes more heavily. Normalize your note format first, or explicitly tell the AI about the quality variation in your input.
Troubleshooting
The AI produces themes that are too generic — labels like 'communication' or 'alignment' with no specificity
Add a specificity rule directly in the prompt: 'Each theme must be stated as a specific, actionable insight — not a topic label. Bad example: Communication issues. Good example: Engineering and CS define done differently, causing repeated re-escalations at handoff.' You can also paste 1–2 example themes from a past synthesis to anchor the AI's output level.
The AI ignores disagreements and produces only consensus findings
Move the disagreements section to the top of your numbered list — not buried at position 3. AI models give more weight to instructions that appear early. Add the phrase: 'Treat surfacing disagreements as equally important as identifying themes. If you find no genuine disagreements, say so explicitly and explain why.' This prevents the AI from smoothing over conflicts to produce a tidier narrative.
The output is too long and includes narrative paragraphs instead of scannable sections
Add an explicit length and format constraint: 'Total output must fit on one printed page — approximately 500 words. Use bullet points and short sentences throughout. Do not write narrative paragraphs for any section.' If the AI continues to overwrite, paste a length-compliant example output and say: 'Match this format and length exactly.'
Next actions are vague and non-committal — 'explore options' or 'gather more input'
Tighten the action instruction: 'Each next action must begin with a specific verb and describe a concrete deliverable — not a process. Bad: Gather more input from CS. Good: CS lead to run 3 follow-up interviews with churned accounts by end of sprint and share findings in the Slack channel.' Adding a deliverable requirement forces the AI away from process language and toward accountable outputs.
The AI attributes themes to the wrong stakeholder roles or conflates multiple roles
Ensure your source headers are consistent and unambiguous before pasting. If roles overlap — for example, two interviewees are both 'Senior PM' — differentiate them with a qualifier like 'Senior PM, Growth' vs. 'Senior PM, Platform.' Also add the instruction: 'When attributing a theme, cite only the roles that explicitly raised it. Do not attribute a theme to a role based on inference.'
How to measure success
How to Evaluate the Quality of Your Synthesis Output
Before you share a readout with any stakeholder, run it through this checklist.
Themes:
- Each theme is stated as an insight, not a topic label
- Every theme includes at least one direct quote with a role attribution
- The number of themes matches what you specified in the prompt
- No two themes are functionally identical or heavily overlapping
Disagreements:
- Conflicts name the specific roles in tension, not vague "some stakeholders"
- Root cause hypotheses are plausible given what you know about your org
- The disagreements section does not simply restate themes in conflict language
Next actions:
- Every action begins with a specific verb and names a concrete deliverable
- Owner types match your actual team structure
- Effort and impact ratings feel calibrated — not everything should be High impact
Accuracy:
- Every direct quote appears verbatim in your source notes
- No theme rests on a single interview without being flagged as low-confidence
- Any gap or missing data is surfaced as a question, not paraphrased around
Now try it on something of your own
Reading about the framework is one thing. Watching it sharpen your own prompt is another — takes 90 seconds, no signup.
Turn 12 stakeholder interviews into a VP-ready one-pager with themes, conflicts, and next actions — before your planning session tomorrow.
Try one of these
Frequently asked questions
There's no strict minimum, but this prompt works best with notes from at least 5–6 interviews. Fewer interviews may not generate reliable themes or meaningful disagreements. For very short note sets (2–3 interviews), ask the AI to identify observations rather than themes and to flag low-confidence patterns explicitly. Most AI models handle up to 8,000–10,000 words of pasted notes comfortably in a single session.
Replace names with role labels before pasting — for example, 'VP of Engineering' instead of 'Sarah Chen.' This protects confidentiality and actually improves the output quality, because the AI attributes themes to roles rather than individuals. Role-based attribution is more useful in readouts, since it signals organizational weight rather than personal opinion. You can do a quick find-and-replace in any text editor before running the prompt.
Yes, but label your source types clearly in the notes. Add a header like 'Interview transcript — CS Manager' or 'Survey response — Enterprise segment' before each block. This lets the AI attribute themes accurately and distinguish between qualitative depth and survey breadth. Without source labels, the AI may conflate a single strong interview voice with broad stakeholder consensus.
This happens occasionally, especially with long note sets. The prompt's accuracy constraint — 'don't invent details; flag missing info as questions' — reduces this risk, but always cross-check any direct quotes against your source notes before sharing a readout. If fabrication is frequent, add the instruction: 'Use only verbatim text from the notes for quotes. If no suitable quote exists for a theme, say so explicitly rather than paraphrasing.'
Replace the owner types and rating scales in the prompt to match your team's language. For example, swap PM/Eng/CS for Design/Data/Growth if that reflects your org. Change S/M/L effort to story points, sprint counts, or quarter-based timeframes if your team uses those. The AI follows whatever tagging convention you specify — just be consistent within the prompt so the output is uniform.
Add a specificity instruction to the prompt: 'Each theme must be a specific, falsifiable insight — not a generic category. Avoid themes like 'communication issues.' Instead, state the specific gap, pattern, or belief.' You can also add 2–3 example themes from a past synthesis as a format anchor. Few-shot examples are the fastest way to push the AI toward the granularity you need.
Light cleanup helps — remove filler words, redundant repetition, and any formatting that could confuse the AI (like nested tables or special characters). But don't rewrite or reinterpret your notes before pasting, since you may inadvertently edit out the nuance the AI needs to find real conflicts. Remove names, label sources by role, and correct obvious typos. Everything else can stay raw.
Yes, and it's a useful technique. Run the prompt twice with slightly different section priorities — for example, once optimized for risk surfacing and once for opportunity identification. Comparing outputs helps you spot gaps and build a more complete synthesis. You can also ask the AI to critique its first output and identify what it may have missed before generating a second pass.