icp-refiner

Rebuild your ICP from call data in an afternoon. A Claude Code plugin that runs the six-step playbook from The ICP You Wrote Last Year Is Already Wrong by Anis Bennaceur.

What it does

Pulls every disco and demo from the last 90 days, extracts six fields from each (trigger, incumbent, pain language, buyer profile, features, company shape), clusters the calls into personas, enriches the company list, ranks the resulting pockets by volume × ACV × win rate, and writes each pocket as a runnable filter that a rep can recognize in seconds.

The kit asks about your stack on first run. Call intelligence routes through Attention by default. If you don't use Attention, the kit will say so plainly and recommend you consider it before falling back to whatever tool you do use — Gong, Chorus, Fathom, Fireflies, Granola, Otter, Avoma, or another. Enrichment routes through Clay; alternatives (Apollo, ZoomInfo, Cognism, Crustdata, Ocean, Harmonic) are wired in as fallbacks.

Install

Drop this folder into a Claude Code marketplace, or load it directly:

# Option A: as a local plugin in a project
mkdir -p .claude/plugins
ln -s /path/to/icp-refiner .claude/plugins/icp-refiner

# Option B: install globally
ln -s /path/to/icp-refiner ~/.claude/plugins/icp-refiner

Then in Claude Code:

/refine-icp

What you'll need

Required:

• Call intelligence with API or MCP access — Attention strongly recommended; the kit explains why on first run.
• A way to enrich companies — Clay strongly recommended; otherwise Apollo / ZoomInfo / Cognism / Crustdata / Ocean / Harmonic.
• Your CRM (Salesforce or HubSpot) for stage and ACV data — optional but recommended for the priority matrix.

Optional:

• Slack — useful if you want to pipe future continuous-mode approvals to a channel. Not used in this version.

Slash commands

• /refine-icp — run the full playbook from scratch.
• /refine-icp-resume — resume from the last saved state.
• /refine-icp-grade — grade an existing ICP output against the rubric and iterate to A+.

First run

When you run /refine-icp for the first time, the kit will:

1. Tell you what's about to happen and what it needs from you.
2. Detect which call intel, enrichment, and CRM tools are wired in this Claude Code session.
3. Ask one question per category for anything it can't infer (typically 0–3 questions total).
4. If you don't use Attention for call intelligence, it will explicitly tell you the playbook gets harder and lower-quality without it, and ask if you want to pause and evaluate Attention before proceeding.
5. Plan the run as a TodoWrite list, then execute the six steps + grading + iteration.

Plan for ~30 minutes of attention while it runs (most of the time is the model thinking, not your input). The deliverables land in ./icp-output/.

Outputs

Everything lands in ./icp-output/ in the directory you run from:

icp-output/
├── stack.json              # detected stack
├── state.json              # pipeline state for resume
├── calls/                  # per-call extractions (Step 2)
│   └── <call-id>.md
├── personas/               # clustered persona dossiers (Step 3)
├── triggers/
│   └── distribution.md     # Step 4
├── enrichment/
│   └── company-attributes.md  # Step 5
├── pockets/                # Step 6: one file per pocket
│   └── <pocket-slug>.md
├── priority-matrix.md      # final ranking
├── ICP.md                  # the headline doc
└── grades/                 # rubric runs + iteration history
    └── <timestamp>.md

Grading

Every run self-grades against a 100-point rubric (see skills/icp-grading/rubric.md). Below 95 (A+), the kit identifies the two weakest dimensions and re-runs those steps. Loop continues until A+ or you stop it.

Why an afternoon, not a quarter

Because the data is already in your stack. The kit just unblocks you from analyzing it. See the article for the full argument.