customer-discovery

A Claude Code skill for working with customer-discovery / user-interview transcripts. Two modes:

• /customer-discovery update — harvests new Gemini-generated meeting notes from your Google Drive, classifies them, and splits each call into three artefacts: full transcript, unbiased pre-pivot transcript (genuine discovery), and post-pitch feedback transcript (reactions to your pitch).
• /customer-discovery <question> — answers ad-hoc questions of the corpus with cited quotes. No infrastructure beyond grep + a subagent.

The pre-pitch / post-pitch split matters: insight extraction shouldn't train on the biased half, but pitch-response analysis needs only the biased half.

Prerequisites

• Claude Code installed.
• Google Drive MCP server connected and authenticated. Verify with /mcp in Claude Code.
• Python 3 (stdlib only — the tooling has zero pip dependencies).

Your Google OAuth credential lives entirely inside the Drive MCP server's own config. No API keys or tokens are stored in this skill. The skill is read-only against your Drive.

Install

Clone this repo and copy the files into the repository where you want your discovery transcripts to live:

git clone https://github.com/<you>/customer-discovery /tmp/cd

cd /path/to/your/repo
mkdir -p .claude/skills/customer-discovery discovery/tooling

cp /tmp/cd/SKILL.md .claude/skills/customer-discovery/
cp /tmp/cd/tooling/*.py discovery/tooling/
cp /tmp/cd/tooling/config.example.json discovery/tooling/config.json

Now run /customer-discovery in Claude Code. On first invocation the skill notices the unconfigured state, asks you for your first name and a Drive folder URL, and writes the real config for you. You won't need to touch any JSON. Once setup wraps, it kicks off the harvest.

Usage

Bootstrap your corpus:

/customer-discovery update

Ask questions:

/customer-discovery what pain points have CPOs mentioned around competitive intel?
/customer-discovery how did people react to my pitch about agentic discovery insights?
/customer-discovery --both has anyone mentioned PMM tooling?

Scope auto-routes from the question (pain words → unbiased; reaction words → feedback; otherwise both), with --pain / --pitch / --both overrides.

What gets stored locally

discovery/                              # name configurable via output_dir
├── LOG.md                              # one row per call: classification, pivot offset, paths
├── full raw transcripts/<slug>.md      # full transcript
├── unbiased raw transcripts/<slug>.md  # pre-pitch portion
├── feedback raw transcripts/<slug>.md  # post-pitch portion (only when a pivot is detected)
└── tooling/                            # config + scripts (this repo)
    ├── config.json                     # your edited config (gitignored if you copy from .example)
    ├── config.py                       # config loader
    ├── parse_transcript.py             # classify + extract + split
    ├── process_batch.py                # batch driver
    └── harvest_dumps.py                # MCP dump → raw/<fileId>.json

Schedule a daily refresh (optional)

If you have the Anthropic Cowork schedule skill installed:

/schedule customer-discovery update

…and pick a cron expression. The routine runs as a remote Claude Code agent, executes /customer-discovery update, and commits the results to your branch. Wake up to a fresh LOG.md.

Tuning to your voice

The defaults are calibrated to one specific interviewer's verbal tics ("let me share what I've heard from other CPOs", "I'm gonna switch to a demo", etc). They'll catch the obvious cases out-of-the-box, but you'll get cleaner pivot cuts after a few runs.

If you notice the pivot cutting too early or too late on your calls, just tell the skill: "the pivot is cutting too early on my Brad calls" or "it's missing my demo opener phrasing". The skill will adjust pivot.patterns and rewrite the config for you — no regex editing required.

Same for the classifier ("it's labelling internal syncs as discovery — tighten up").

License

MIT — see LICENSE.