Product Discovery Skills

A pipeline of Claude Code skills that take a folder of customer interview transcripts and turn it into a prioritized list of opportunities, end-to-end. Rooted in Teresa Torres' Continuous Discovery Habits and the Opportunity Solution Tree (OST).

The skills work together as a pipeline, but each is independently invocable — install one, several, or all of them.

Pipeline

icp-screener → phase-map-analyst → opportunity-analyst → opportunity-clusterer → opportunity-sizer
                       ↑                    │
                       └───── feedback ─────┘
                       (via opportunity-analyst-reset)

flowchart LR
  A[icp-screener] --> B[phase-map-analyst]
  B --> C[opportunity-analyst]
  C --> D[opportunity-clusterer]
  D --> E[opportunity-sizer]
  C -. misfits .-> B
  R[opportunity-analyst-reset] -. cleanup before re-run .-> C

Skills

#	Skill	What it does	Status
1	`icp-screener`	Filter raw transcripts to ICP matches; produce a TEMP folder	shipped
2	`phase-map-analyst`	Build the first level of the OST: 3-8 phases of the JTBD	shipped
3	`opportunity-analyst`	Extract opportunities per transcript; tag each to a phase	shipped
4	`opportunity-analyst-reset`	Strip prior analyst output so the pipeline can re-run cleanly	shipped
5	`opportunity-clusterer`	Cluster opportunities across transcripts; hard-partition by phase; verbatim labels	shipped
6	`opportunity-sizer`	Score and prioritize clusters on importance × prevalence; surface the top focus	shipped

Install

Quick install (all skills):

git clone https://github.com/Elsevanderberg1/product-discovery-skills.git
cd product-discovery-skills
./install.sh

Install only specific skills:

./install.sh icp-screener
./install.sh phase-map-analyst opportunity-analyst opportunity-analyst-reset

Skills install to ~/.claude/skills/<skill-name>/. Re-running install with the same skill replaces it.

Quick start

After installing, point any of the skills at a folder of .md interview transcripts:

/icp-screener /path/to/transcripts

This produces an icp-screened-TEMP-YYYY-MM-DD/ subfolder. Then run the rest of the pipeline against that folder:

/phase-map-analyst /path/to/transcripts/icp-screened-TEMP-YYYY-MM-DD
/opportunity-analyst /path/to/transcripts/icp-screened-TEMP-YYYY-MM-DD
/opportunity-clusterer /path/to/transcripts/icp-screened-TEMP-YYYY-MM-DD
/opportunity-sizer /path/to/transcripts/icp-screened-TEMP-YYYY-MM-DD

How the skills relate

Each skill consumes the previous skill's TEMP folder. The artifacts inside that folder (the screening overview, the phase-map markdown, appended sections in transcripts) are the data contract between skills. If you change one skill's output format, the consuming skill needs to know — that's why this is a single repo with one fork-point and one versioned history.

The feedback loop matters: phase-map-analyst can detect existing opportunity-analyst output in the TEMP folder and offer to revise the map based on the misfit evidence. After revision, opportunity-analyst-reset strips the stale tagging so opportunity-analyst can re-run cleanly against the new map.

Concepts

Job-to-be-Done (JTBD): the functional job the customer is trying to get done.
Phase map: a 3-8 step decomposition of the JTBD; the first level of the Opportunity Solution Tree.
Opportunity: a pain, friction, wish, or desire — framed in the customer's voice, tagged to a phase.
Importance × Prevalence: importance = how much one customer cares (1-5 scale, inferred only when stated). Prevalence = how many customers express the opportunity. Combined, they're the sizer's ranking signal.
Misfit bucket: opportunities that move the product metric but don't fit any phase. Treated as evidence the phase map may need revision, not as noise to discard.

For deeper context, read Torres' Continuous Discovery Habits and her writing on the Opportunity Solution Tree.

Design principles

Skills, not pipelines. Each skill is independently invocable. The pipeline is a convention, not a runtime.
Markdown all the way down. Inputs, outputs, and artifacts are all .md files. Human-readable, diff-friendly, no special tooling required.
Read-mostly on transcripts. Only opportunity-analyst and its reset modify transcript files (by appending or stripping a clearly-labelled section). Everything else is read-only.
Misfits are signal. Opportunities that don't fit any phase are surfaced explicitly, not silently re-tagged or dropped.

Contributing

Issues and PRs welcome. If you propose a new skill or a change to an existing skill's data contract (i.e. anything another skill in the pipeline reads), please open an issue first to discuss the inter-skill consequences.

License

MIT.

Author

Else van der Berg — freelance product lead working with AI-forward startups and scale-ups.