C4PM

Cursor for Product Management — a CLI that turns customer interview transcripts into ranked problems and agent-ready product specs.

Drop a folder of interview transcripts. Get a ranked list of what to build, backed by exact user quotes. Feed the spec to Cursor or Claude Code and start building.

c4pm analyze interviews/
c4pm spec interviews/ -o spec.json
cursor spec.json   # or feed to any coding agent

---

How It Works

transcripts/          c4pm analyze        c4pm spec
┌──────────┐         ┌──────────────┐    ┌──────────────┐
│ .txt/.md  │───────▶│ Extract      │───▶│ JSON spec    │───▶ Cursor / Claude Code
│ files     │        │ Rank         │    │ ready to     │     starts building
└──────────┘         │ Score (1-16) │    │ implement    │
                     └──────────────┘    └──────────────┘

Three AI steps, zero manual work:

1. Extract — Clusters raw feedback into 4-7 distinct problems with exact quotes and speaker attribution
2. Rank — Scores each problem on Reach, Intensity, User Value, and Confidence (max 16 points), using reasoning effort for comparative judgment
3. Spec — Generates a structured JSON spec for the #1 problem with user stories, acceptance criteria, API design, and evidence

---

Install

git clone https://github.com/your-org/c4pm.git
cd c4pm
pip install -e .

Set your OpenAI API key:

export OPENAI_API_KEY=sk-...
# or create a .env file
echo "OPENAI_API_KEY=sk-..." > .env

Requires Python 3.9+.

---

Usage

Analyze: Find and rank problems

c4pm analyze <transcript-dir> [OPTIONS]

Option	Description
--top, -n	Number of problems to show (default: 5)
--verbose, -v	Show detailed extraction and ranking logs

Example:

$ c4pm analyze sample_data/ -n 3

╭──────────────────────────────────────────────────────╮
│ C4PM - Analyzing customer feedback                   │
╰──────────────────────────────────────────────────────╯

Loading transcripts...
  Loaded 6 transcripts
    - Sarah Chen (Engineering Manager)
    - Marcus Johnson (Head of Product)
    - Priya Patel (Founder/CEO)
    - David Kim (Senior Product Manager)
    - Meera Krishnamurthy (Head of Product)
    - Raj Patel (Senior Product Manager)

Extracting problems...
  Found 5 distinct problems

Ranking by impact...

============================================================
PROBLEM RANKING SUMMARY

┏━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ #   ┃ Problem                      ┃ Score ┃ Severity   ┃ Confidence ┃
┡━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ 1   │ Manual, Biased Feedback      │ 15/16 │ major_pain │ high       │
│     │ Synthesis                    │       │            │            │
│ 2   │ No Evidence-Based            │ 14/16 │ major_pain │ high       │
│     │ Prioritization Framework     │       │            │            │
│ 3   │ Lack of Automated Problem    │ 13/16 │ major_pain │ high       │
│     │ Clustering                   │       │            │            │
└─────┴──────────────────────────────┴───────┴────────────┴────────────┘

Each problem includes:

• Scoring breakdown — Reach (1-5), Intensity (1-5), User Value (1-3), Confidence (1-3)
• Evidence — Up to 3 exact quotes with speaker name and role
• Mentioned by — Which interviewees raised the issue
• Urgency signals — Emotional language detected (e.g. "drowning", "broken", "disaster")
• Reasoning — Why this problem ranks where it does
• Tradeoffs — What happens if you don't solve it

---

Spec: Generate a build spec

c4pm spec <transcript-dir> [OPTIONS]

Option	Description
--output, -o	Save spec to a JSON file (default: stdout)
--verbose, -v	Show detailed logs

Example:

$ c4pm spec sample_data/ -o spec.json

The generated spec is structured JSON containing:

problem_statement        What exactly is broken, for whom, with evidence
user_stories             5 role-specific stories with action and benefit
proposed_solution        Summary, key features, data model, API changes, workflow
acceptance_criteria      Testable conditions for done
out_of_scope             What NOT to build
success_metrics          Leading and lagging indicators
risks                    Technical, adoption, integration risks with mitigations
implementation_hints     Concrete tech stack and architecture suggestions
evidence_summary         Direct user quotes backing the spec
priority_justification   Why this problem was ranked #1

Feed the JSON directly to a coding agent:

# Cursor
cursor spec.json

# Claude Code
cat spec.json | claude "implement this spec"

---

Transcript Format

Each interview is a .txt or .md file. Add metadata in the first few lines for better attribution:

Interviewee: Sarah Chen
Role: Engineering Manager
Company: Acme Corp
Date: 2024-01-15

---

Interviewer: Tell me about your current workflow...

Sarah: Honestly, it's a mess. We have user feedback
coming from everywhere...

Supported metadata fields: Interviewee, Role, Company, Date, User Type

All fields are optional. C4PM works without metadata — you just get less precise attribution.

---

Scoring Framework

Each problem is scored across four dimensions:

Dimension	Range	What it measures
Reach	1-5	How many users are affected
Intensity	1-5	How painful the problem is
User Value	1-3	How valuable the affected users are
Confidence	1-3	How strong the evidence is

Total: /16 points. No ties — every problem gets a unique score with head-to-head comparison.

Hard rules enforce honesty:

• If only 1 user mentioned it → Confidence = 1
• If fewer than 3 interviews → Confidence capped at 2

---

Project Structure

c4pm/
  cli.py              CLI entry point (Typer + Rich)
  llm.py              Shared OpenAI client, retry/backoff, JSON parsing
  ingest/
    loader.py          Load .txt/.md transcripts, extract metadata
  reasoning/
    extractor.py       Cluster feedback into problems (gpt-5.4-nano)
    ranker.py          Score and rank problems (gpt-5.4-nano, reasoning effort)
  output/
    spec.py            Generate structured JSON spec (gpt-5.4-nano)

---

Requirements

• Python 3.9+
• OpenAI API key
• Dependencies: typer, rich, openai, python-dotenv

---

License

MIT