Research workflow

Customer feedback to themes and roadmap asks

Tags pasted reviews, NPS comments or tickets in parallel chunks, clusters them into counted themes, backs each theme with verbatim quotes, optionally compares competitors' public reviews, and writes a ranked report with roadmap asks framed as problems.

ParallelPrompt chainEvaluator–optimizerHuman approvalRuns in Claude Code, Claude.ai, ChatGPT

Running it here

Every step runs here with Claude when you're signed in, and in your own agent.

  1. TriggerPaste the feedbackPaste {{feedback}} (reviews, NPS comments, survey answers or a ticket export) with {{product_context}}, your {{roadmap}} and optional {{competitors}}. Remove customer emails and names first if the report will be shared.→ manual
  2. PromptNumber and clean the rowsTurn {{feedback}} into a numbered list R1, R2…, one row per piece of feedback. Keep the exact text: do not shorten, translate or fix spelling. Keep any metadata on the row (score, plan, date, channel, segment) as key: value. Drop empty rows and exact duplicates. End with: rows kept, rows dropped (and why), which metadata fields exist, and the rows split into four even ranges labelled Chunk 1 to Chunk 4 (e.g. Chunk 1 = R1–R50). If there are fewer than 20 rows, put them all in Chunk 1 and mark Chunks 2–4 empty.
  3. ParallelTag rows in parallelTag the rows in {{rows}} that fall in your chunk (the ranges are listed at its end). For each row write one line: ID | sentiment (positive, mixed, negative) | theme (a 2–4 word noun phrase naming the underlying need, e.g. 'slow CSV export', not 'bad') | feature (an area from {{product_context}}, or 'general') | segment (only from the row's metadata, else 'unknown') | severity (1 = cosmetic, 2 = annoyance, 3 = blocks a task, 4 = threatens churn or money; 'n/a' for praise). A row may have up to 2 themes. Tag only what the row says. If your chunk is empty, write 'No rows in this chunk.'
    • Chunk 1Tag every row in Chunk 1.
    • Chunk 2Tag every row in Chunk 2.
    • Chunk 3Tag every row in Chunk 3.
    • Chunk 4Tag every row in Chunk 4.
  4. AgentPublic reviews (optional)If {{competitors}} is empty or says 'none', reply 'Skipped: no competitors given.' and stop. Otherwise, for each competitor, search for reviews from the last 12 months on G2, Capterra, Trustpilot, the App Store, Google Play and Reddit. Some review sites block automated reading: list which ones you couldn't open and use only pages you could read. Report the 3 most common complaints and 2 most common praises you saw, each with 1–2 short verbatim quotes and the URL. Then say which of them overlap with themes in {{tags}}. Do not estimate counts you didn't see; say how many reviews you actually read.
  5. PromptCluster into themesMerge the theme labels in {{tags}} into 6–12 themes. Combine labels that describe the same need; keep a theme separate if fixing it would take different work. For each theme give: name, one-sentence definition, row count (a row counts once per theme), share of all rows, sentiment split, average severity (ignoring 'n/a'; write 'praise' if all rows are n/a), top segments, main feature, and the full list of row IDs. Put themes with fewer than 2 rows into 'Long tail' with their IDs. Finally, state how many rows have at least one theme and confirm it matches the count in {{rows}}; if it doesn't, say which IDs are missing.
  6. PromptVerify with quotesFor each theme in {{themes}}: pick 2–3 rows from its ID list and copy their text from {{rows}} word for word, with the ID. Prefer rows from different segments. Then re-read 5 IDs from the theme's list and check they fit the definition; if more than 1 of the 5 doesn't fit, re-check the whole list and correct the count. Drop any theme without 2 supporting rows and say why. Return the verified theme table (same columns as before, corrected counts) and the quotes under each theme, then a 'Changes' list.
  7. PromptWrite insights and roadmap asksFrom {{verified}}, and {{public}} if it has data, write the report for {{product_context}}: 1. Summary: 5 bullets, each a finding with a number. 2. Problem theme table sorted by opportunity score = row count × average severity (praise themes go in section 5, not here): Theme | Rows | Share | Avg severity | Score | Segments | Trend (only if dates exist). 3. Top 5 insights: what users struggle with, who, how badly, and evidence (count, row IDs, one quote). 4. Roadmap asks for the top 5 problem themes: a problem statement (not a solution), the evidence, a success measure, and status against {{roadmap}} (on roadmap, partly, or new). 5. What's working: the top 3 praise themes. 6. Market context from public reviews, if any, with URLs. 7. Caveats: sample size, source bias (e.g. reviews skew negative), segments with too few rows to conclude, and how many rows were 'unknown' segment. Use only numbers from {{verified}}.
  8. Review loopCheck the report1. Every insight and roadmap ask states a row count and at least one row ID (R-number). 2. Every quote is in quotation marks and followed by a row ID. 3. Every opportunity score equals rows × average severity (recompute each one). 4. Roadmap asks are written as problems ('users can't export a full year') not solutions ('build async export'). 5. A Caveats section exists and states the sample size. 6. No claims about causes, revenue or churn rates that the counts can't show. 7. Under 1,200 words, not counting tables.
  9. ApprovalApprove the reportReview {{report}} and spot-check a few quotes in {{verified}} against your source data. Approve, edit or reject before sharing.→ chat
  10. OutputSave the report and theme tableReturn {{report}} and {{verified}}.→ file

Good for

Product and support teams turning a pile of reviews, NPS comments or tickets into counted themes.

Last reviewed

Install as a skill

The whole workflow as one SKILL.md. Install it in Claude Code, Codex or Copilot and ask your agent to run “Customer feedback to themes and roadmap asks”.

Get the customer-feedback-theme-report skill

What each step does

  1. 1. Paste the feedback (Trigger)

    Paste {{feedback}} (reviews, NPS comments, survey answers or a ticket export) with {{product_context}}, your {{roadmap}} and optional {{competitors}}. Remove customer emails and names first if the report will be shared.

  2. 2. Number and clean the rows (Prompt)

    Turn {{feedback}} into a numbered list R1, R2…, one row per piece of feedback. Keep the exact text: do not shorten, translate or fix spelling. Keep any metadata on the row (score, plan, date, channel, segment) as key: value. Drop empty rows and exact duplicates. End with: rows kept, rows dropped (and why), which metadata fields exist, and the rows split into four even ranges labelled Chunk 1 to Chunk 4 (e.g. Chunk 1 = R1–R50). If there are fewer than 20 rows, put them all in Chunk 1 and mark Chunks 2–4 empty.

  3. 3. Tag rows in parallel (Parallel)

    Tag the rows in {{rows}} that fall in your chunk (the ranges are listed at its end). For each row write one line: ID | sentiment (positive, mixed, negative) | theme (a 2–4 word noun phrase naming the underlying need, e.g. 'slow CSV export', not 'bad') | feature (an area from {{product_context}}, or 'general') | segment (only from the row's metadata, else 'unknown') | severity (1 = cosmetic, 2 = annoyance, 3 = blocks a task, 4 = threatens churn or money; 'n/a' for praise). A row may have up to 2 themes. Tag only what the row says. If your chunk is empty, write 'No rows in this chunk.'

    • Chunk 1: Tag every row in Chunk 1.
    • Chunk 2: Tag every row in Chunk 2.
    • Chunk 3: Tag every row in Chunk 3.
    • Chunk 4: Tag every row in Chunk 4.
  4. 4. Public reviews (optional) (Agent)

    If {{competitors}} is empty or says 'none', reply 'Skipped: no competitors given.' and stop. Otherwise, for each competitor, search for reviews from the last 12 months on G2, Capterra, Trustpilot, the App Store, Google Play and Reddit. Some review sites block automated reading: list which ones you couldn't open and use only pages you could read. Report the 3 most common complaints and 2 most common praises you saw, each with 1–2 short verbatim quotes and the URL. Then say which of them overlap with themes in {{tags}}. Do not estimate counts you didn't see; say how many reviews you actually read.

  5. 5. Cluster into themes (Prompt)

    Merge the theme labels in {{tags}} into 6–12 themes. Combine labels that describe the same need; keep a theme separate if fixing it would take different work. For each theme give: name, one-sentence definition, row count (a row counts once per theme), share of all rows, sentiment split, average severity (ignoring 'n/a'; write 'praise' if all rows are n/a), top segments, main feature, and the full list of row IDs. Put themes with fewer than 2 rows into 'Long tail' with their IDs. Finally, state how many rows have at least one theme and confirm it matches the count in {{rows}}; if it doesn't, say which IDs are missing.

  6. 6. Verify with quotes (Prompt)

    For each theme in {{themes}}: pick 2–3 rows from its ID list and copy their text from {{rows}} word for word, with the ID. Prefer rows from different segments. Then re-read 5 IDs from the theme's list and check they fit the definition; if more than 1 of the 5 doesn't fit, re-check the whole list and correct the count. Drop any theme without 2 supporting rows and say why. Return the verified theme table (same columns as before, corrected counts) and the quotes under each theme, then a 'Changes' list.

  7. 7. Write insights and roadmap asks (Prompt)

    From {{verified}}, and {{public}} if it has data, write the report for {{product_context}}: 1. Summary: 5 bullets, each a finding with a number. 2. Problem theme table sorted by opportunity score = row count × average severity (praise themes go in section 5, not here): Theme | Rows | Share | Avg severity | Score | Segments | Trend (only if dates exist). 3. Top 5 insights: what users struggle with, who, how badly, and evidence (count, row IDs, one quote). 4. Roadmap asks for the top 5 problem themes: a problem statement (not a solution), the evidence, a success measure, and status against {{roadmap}} (on roadmap, partly, or new). 5. What's working: the top 3 praise themes. 6. Market context from public reviews, if any, with URLs. 7. Caveats: sample size, source bias (e.g. reviews skew negative), segments with too few rows to conclude, and how many rows were 'unknown' segment. Use only numbers from {{verified}}.

  8. 8. Check the report (Review loop)

    1. Every insight and roadmap ask states a row count and at least one row ID (R-number). 2. Every quote is in quotation marks and followed by a row ID. 3. Every opportunity score equals rows × average severity (recompute each one). 4. Roadmap asks are written as problems ('users can't export a full year') not solutions ('build async export'). 5. A Caveats section exists and states the sample size. 6. No claims about causes, revenue or churn rates that the counts can't show. 7. Under 1,200 words, not counting tables.

  9. 9. Approve the report (Approval)

    Review {{report}} and spot-check a few quotes in {{verified}} against your source data. Approve, edit or reject before sharing.

  10. 10. Save the report and theme table (Output)

    Return {{report}} and {{verified}}.

Use it in your own agent

Fill in the inputs
Too long for an Open in link: copy the chat prompt below instead.4 inputs use their example until you fill them in.

Paste into ChatGPT, Claude or Gemini. Parallel steps run one after another, and approvals become “wait for my OK”.

customer-feedback-theme-report.txt
You're going to run a multi-step workflow: Customer feedback to themes and roadmap asks.

Inputs:
- feedback: Paste one piece of feedback per line, with any metadata, e.g. 'NPS 3 | plan: Pro | 2026-10-04 | Export to CSV takes minutes and times out on big months'
- product_context: Ledgerly, invoicing for freelancers. Areas: invoices, reminders, payments, export, mobile app, onboarding.
- roadmap: Q1: faster export, Stripe payouts. Q2: mobile app redesign. (or 'none')
- competitors: FreshBooks, Wave (or 'none' to skip public reviews)

Work through these steps in order. Label each step's output with its number. Where a step says to wait for my OK, stop and wait.

1. **Number and clean the rows** — Turn Paste one piece of feedback per line, with any metadata, e.g. 'NPS 3 | plan: Pro | 2026-10-04 | Export to CSV takes minutes and times out on big months' into a numbered list R1, R2…, one row per piece of feedback. Keep the exact text: do not shorten, translate or fix spelling. Keep any metadata on the row (score, plan, date, channel, segment) as key: value. Drop empty rows and exact duplicates. End with: rows kept, rows dropped (and why), which metadata fields exist, and the rows split into four even ranges labelled Chunk 1 to Chunk 4 (e.g. Chunk 1 = R1–R50). If there are fewer than 20 rows, put them all in Chunk 1 and mark Chunks 2–4 empty. Save the result as `rows`.
2. **Tag rows in parallel** — Do each of these separately, then combine them. Tag the rows in {{rows}} that fall in your chunk (the ranges are listed at its end). For each row write one line: ID | sentiment (positive, mixed, negative) | theme (a 2–4 word noun phrase naming the underlying need, e.g. 'slow CSV export', not 'bad') | feature (an area from Ledgerly, invoicing for freelancers. Areas: invoices, reminders, payments, export, mobile app, onboarding., or 'general') | segment (only from the row's metadata, else 'unknown') | severity (1 = cosmetic, 2 = annoyance, 3 = blocks a task, 4 = threatens churn or money; 'n/a' for praise). A row may have up to 2 themes. Tag only what the row says. If your chunk is empty, write 'No rows in this chunk.' Save the result as `tags`.
   - **Chunk 1**: Tag every row in Chunk 1.
   - **Chunk 2**: Tag every row in Chunk 2.
   - **Chunk 3**: Tag every row in Chunk 3.
   - **Chunk 4**: Tag every row in Chunk 4.
3. **Public reviews (optional)** — If FreshBooks, Wave (or 'none' to skip public reviews) is empty or says 'none', reply 'Skipped: no competitors given.' and stop. Otherwise, for each competitor, search for reviews from the last 12 months on G2, Capterra, Trustpilot, the App Store, Google Play and Reddit. Some review sites block automated reading: list which ones you couldn't open and use only pages you could read. Report the 3 most common complaints and 2 most common praises you saw, each with 1–2 short verbatim quotes and the URL. Then say which of them overlap with themes in {{tags}}. Do not estimate counts you didn't see; say how many reviews you actually read. Save the result as `public`.
4. **Cluster into themes** — Merge the theme labels in {{tags}} into 6–12 themes. Combine labels that describe the same need; keep a theme separate if fixing it would take different work. For each theme give: name, one-sentence definition, row count (a row counts once per theme), share of all rows, sentiment split, average severity (ignoring 'n/a'; write 'praise' if all rows are n/a), top segments, main feature, and the full list of row IDs. Put themes with fewer than 2 rows into 'Long tail' with their IDs. Finally, state how many rows have at least one theme and confirm it matches the count in {{rows}}; if it doesn't, say which IDs are missing. Save the result as `themes`.
5. **Verify with quotes** — For each theme in {{themes}}: pick 2–3 rows from its ID list and copy their text from {{rows}} word for word, with the ID. Prefer rows from different segments. Then re-read 5 IDs from the theme's list and check they fit the definition; if more than 1 of the 5 doesn't fit, re-check the whole list and correct the count. Drop any theme without 2 supporting rows and say why. Return the verified theme table (same columns as before, corrected counts) and the quotes under each theme, then a 'Changes' list. Save the result as `verified`.
6. **Write insights and roadmap asks** — From {{verified}}, and {{public}} if it has data, write the report for Ledgerly, invoicing for freelancers. Areas: invoices, reminders, payments, export, mobile app, onboarding.:
1. Summary: 5 bullets, each a finding with a number.
2. Problem theme table sorted by opportunity score = row count × average severity (praise themes go in section 5, not here): Theme | Rows | Share | Avg severity | Score | Segments | Trend (only if dates exist).
3. Top 5 insights: what users struggle with, who, how badly, and evidence (count, row IDs, one quote).
4. Roadmap asks for the top 5 problem themes: a problem statement (not a solution), the evidence, a success measure, and status against Q1: faster export, Stripe payouts. Q2: mobile app redesign. (or 'none') (on roadmap, partly, or new).
5. What's working: the top 3 praise themes.
6. Market context from public reviews, if any, with URLs.
7. Caveats: sample size, source bias (e.g. reviews skew negative), segments with too few rows to conclude, and how many rows were 'unknown' segment.
Use only numbers from {{verified}}. Save the result as `report`.
7. **Check the report** — Review loop (max 3 passes): 1. Every insight and roadmap ask states a row count and at least one row ID (R-number).
2. Every quote is in quotation marks and followed by a row ID.
3. Every opportunity score equals rows × average severity (recompute each one).
4. Roadmap asks are written as problems ('users can't export a full year') not solutions ('build async export').
5. A Caveats section exists and states the sample size.
6. No claims about causes, revenue or churn rates that the counts can't show.
7. Under 1,200 words, not counting tables. Save the result as `report`.
8. **Approve the report** — Show me the result and wait for my OK before the next step: Review {{report}} and spot-check a few quotes in {{verified}} against your source data. Approve, edit or reject before sharing.
9. **Save the report and theme table** — Finish: Return {{report}} and {{verified}}.

Want different steps? Open it in the workflow builder.