Sales · 01 · Your GTM data machineLesson 1 of 4

Your ICP, written down for machines

45 min working time · Weeks 5-6

By the end of this lesson you can

Encode your ICP as a scoring rubric Claude can apply
Build the qualification skill you'll use everywhere

Why your ICP needs to be a file

Walk out of this lesson with two assets in your repo: your ICP as a scored rubric in markdown, and an /icp-qualify skill that applies it identically to one lead or a thousand.

Every team says they know their ICP. Almost none have it written down precisely enough that a machine - or a new hire - could apply it the same way twice. "We sell to agencies, 10-50 people, in the US" sounds specific until you hit the edge cases: is a 9-person agency out? Does a UK agency with a US office count? Is a freelancer collective an agency?

The 2026 shift in outbound is from volume to precision. The 2026 benchmark reports found elite teams now have AI doing roughly 80% of research and sequencing work, and they lead with intent timing, not bigger lists. Precision starts with a rubric: a written, scoreable definition of who you sell to. Every workflow in this track - list building, enrichment, personalization, routing - reads from this one file.

Mine your closed-won deals first

Don't write the rubric from memory. Your best ICP definition is hiding in your closed-won list. Export your last 20-30 won deals from the CRM and let Claude find the pattern.

Export closed-won deals from your CRM as a CSV (company name, size, industry, deal size, sales cycle length, source).
Open a Claude Code session in your sales workspace folder and drop the CSV in.
Ask: "Analyze these closed-won deals. What do the fastest-closing, highest-value deals have in common? Propose 5-7 scoring dimensions with weights."
Do the same with 10-15 closed-LOST deals. The disqualifiers matter as much as the qualifiers.
Argue with the output. You know things the data doesn't show - regulatory constraints, founder preferences, segments you're exiting. Edit until it matches reality.

The rubric format

Structure the rubric as three weighted groups: firmographic fit (does the company match), signal recency (is something happening right now), and persona weight (is this person the buyer). That grouping comes straight from production GTM systems - it maps cleanly to how you'll score later.

icp-rubric.md (the asset everything else reads)

# ICP Scoring Rubric - v3 (2026-09)

Score each lead 0-100. Disqualifiers zero the score.

## Firmographic fit (50 pts)
- Industry: marketing/growth agency, RevOps consultancy (20)
  - Adjacent (dev shop, design studio): 10
- Headcount 10-50 (15) · 5-9 or 51-100 (8) · else (0)
  (band on the provider's size bucket, never on
   employees_on_linkedin - it undercounts 3-10x)
- Geo: US/UK/CA/AU (10) · EU (5) · else (0)
- Visible revenue motion: outbound or paid case studies (5)

## Signal recency (30 pts)
- Hiring SDR/GTM/RevOps role, posted <60 days (15)
- Funding round closed <90 days (10)
- Leadership change or new service launch <90 days (5)

## Persona weight (20 pts)
- Founder / Managing Partner (20)
- Head of Growth / RevOps lead (15)
- SDR manager (8) · IC (2)
- Match on the ROLE field, whole words only. Never on
  the free-text headline.

## Disqualifiers (score = 0)
- Construction, government, generic local services
- Headcount > 500
- Active customer or open opportunity in CRM
- Peer: sells done-for-you services in our category.
  The line is services vs product - a company selling
  software (even AI software) is a buyer; a build shop
  or agency doing what we do is a competitor.

Keep it under a page. Every number is a decision your team can argue about in a pull request - which is exactly the point. The rubric is editable by non-engineers because it's markdown, not code.

Gates that came from real false positivesThe last three rubric lines above came out of our own signal engine, one embarrassing row at a time. Substring matching on bdr rejected a founder whose headline contained the letters; whole-word matching fixed it. Matching titles against the whole headline instead of the role field let "ex-Head of Sales, now investor" score as a buyer. A minimum of 1,000 followers went in after a senior title on a 200-follower profile turned out to be a fake. An "employer must have a company page" gate cut a 25k-follower founder, so it now carries a founder exemption. And the competitor gate is a judgment, not a keyword: on an early run 11 of 17 "qualified" people were competitors. Every gate you add should be traceable to a row that got through.

Borrow before you build: the skill libraries

Before you build your first skill, learn the move that saves you hours on every skill after it: there are two ways to get a skill - borrow a ready one or build your own. Skills are portable SKILL.md files, which means big free libraries of ready-made ones already exist, and checking them takes two minutes.

SkillsMP (skillsmp.com) - a free independent catalog indexing 3,000,000+ open-source skills from public GitHub as of September 2026, browsable by category, with a free REST API and an MCP server. Search 'enrichment' or 'outbound' and see what someone already wrote.
Anthropic's official skills repo (github.com/anthropics/skills) - document skills (docx, pdf, pptx, xlsx), example skills including skill-creator, and the Agent Skills spec. Install with /plugin marketplace add anthropics/skills, which gives you the document-skills and example-skills plugins.
The official plugin marketplace - /plugin install <name>@claude-plugins-official, or /plugin then Discover to browse. The community marketplace is added with /plugin marketplace add anthropics/claude-plugins-community and installed from as @claude-community.
agentskills.io - the open standard behind the SKILL.md format; the same file also runs in other agent tools that adopted it.
Using a skill from GitHub is manual but trivial: read it, then copy the folder into .claude/skills/. That's the whole install.

A two-million-skill index is mostly noise. Borrow the shape, then read every line before it sits next to your API keys. The split for GTM work: install the generic plumbing - a CSV deduper, an enrichment-API caller, a report formatter - and hand-build only what's yours: your ICP rubric, your voice, your sequence logic, your deliverability defaults. No library knows your business. The qualification skill you build next is the textbook hand-build case, because no library knows your ICP.

Where a skill lives decides who gets it. .claude/skills/ in the project repo is committed and shared with everyone who clones it, which is where the rubric and every GTM skill in this track go. ~/.claude/skills/ is yours on every project and never leaves your machine. A plugin in a marketplace is how a team ships a versioned skill library (the capstone does this). To scaffold a new one, run Anthropic's skill-creator skill; the frontmatter it writes can pin model (so a bulk-classification skill always runs on Haiku), restrict allowed-tools, run in a forked context, or hide the skill from automatic invocation with disable-model-invocation.

Turn it into the qualification skill

A skill is the rubric plus the procedure for applying it. Once it exists, anyone on the team types one command and gets consistent scoring - no prompt crafting, no drift.

.claude/skills/icp-qualify/SKILL.md

---
name: icp-qualify
description: Score leads against our ICP rubric. Use whenever
  leads need qualification - accepts a CSV path or a single
  company/person description.
argument-hint: <csv-path or company description>
model: sonnet
---

You qualify leads for our team.

1. Read the rubric at ./icp-rubric.md - it is the only
   source of truth. Never invent criteria. If the file is
   missing, stop and say so - never score from memory.
2. For a CSV: score every row. Add columns icp_score,
   icp_band (A: 80+, B: 60-79, C: 40-59, D: below 40),
   and icp_reason (one line citing the rubric).
3. For a single lead: return score, band, and a 2-line
   rationale citing specific rubric lines.
4. If a field needed for scoring is missing, score what
   you can and flag the row - never guess values.
5. Apply disqualifiers first. A disqualified lead is 0,
   regardless of other points.
6. Never send, never enrich, never spend a credit. This
   skill reads and scores. Spending is a separate,
   human-approved step.

Create the folder: .claude/skills/icp-qualify/ inside your sales workspace repo.
Save the SKILL.md above, adjusted to point at your rubric file.
Test on 5 leads you know well: /icp-qualify followed by a pasted company description.
Compare against your gut. Where the skill disagrees with you, the rubric is missing a rule - fix the rubric, not the prompt.

Deterministic where possibleFor big lists you'll later move scoring into a script (lesson 4): plain code applies the firmographic math, and Claude is only called for fuzzy judgments like "is this title a decision-maker for our offer?" or "is this company a peer or a buyer?". On our own engine that judgment runs on Haiku 4.5 with a cached system prompt at about $0.02 per 1,000 people. Cheaper, faster, and the same answer every run. The model: sonnet line in the skill above is the interactive default; lesson 4 covers when to pin haiku and when a cheaper open model is honest.

For teams: one rubric, one owner

One file, zero copiesOur signal ICP lives in one YAML file that two different engines read: the one that scores engagers on our own posts and the one that scores engagers on other creators' posts. Neither keeps a local copy, and both refuse to run if the file is missing. A second copy is a guarantee the two will disagree within a month. Every script and skill in this track that needs the ICP reads icp-rubric.md by path; none of them paste it in.

This is the first taste of the track's core move: judgment, written down once, applied by machines forever. Everything that follows builds on this file. It is also the first entry in your company brain: a repo whose CLAUDE.md indexes what exists, where the rubric and voice files live, and which build to copy next time. Keep that index current and every new project starts from your best prior work instead of a blank page.

Do this now

Sources and further reading

We set it up with you

Want us to set it up with you, end to end?

Three one-on-one sessions. We train you on your real stack and build your first agents together, until you can run it yourself. You keep everything.