gtm-icp

gtm-icp

GTM-as-code: a backend-free Claude plugin that finds B2B accounts matching your ICP, then enriches, classifies and scores them into a prioritized target list for outreach — grounded on your own corpus, bring your own keys.

Working name. The pipeline runs end to end — discover → enrich (+ signals) → classify → score → people → list → personalize — from an ICP brief to a ranked, evidence-backed target list with contacts and grounded first-draft outreach.

Most ICP/lead-scoring tools are hosted SaaS that mark up data you already pay for and score accounts with a generic prompt. gtm-icp flips both:

• No hosted backend. The pipeline is a set of Claude skills with bundled, stdlib-only Python scripts. Clone it, install it as a plugin, run it.
• Bring your own keys, no markup. Enrichment calls your Apollo key directly (sync REST). The orchestration and LLM reasoning cost cents; the data is yours at vendor cost, not resold.
• Grounded, not guessed. Classification cites your won-deal patterns, positioning, and case studies — from a local corpus or a K2 corpus — instead of a bare LLM judgment. That grounding is the part a commodity wrapper can't trivially copy.
• A real no-key path. Without any paid key, enrich runs firmographic-only and classify grounds on local corpus files. You get a working clone-and-run demo with zero secrets.

Pipeline stages

Stage	Status	What it does
discover	implemented	Find ICP-fit companies from a brief — Perplexity-grounded, with no-key DuckDuckGo + seed-list fallbacks.
enrich	implemented	Firmographics (Apollo-first, no-key fallback) + deep intent signals from website, public ATS job boards (Greenhouse/Lever/Ashby), and GitHub — mapped to ICP scoring dimensions.
classify	implemented	Corpus-grounded ICP verdict — hard gates + graded dimensions → deterministic 0-100 score & A/B/Nurture/Reject tier.
people	implemented	Apollo people search — map the ICP's buying personas to real contacts per qualified account; no-key persona-target fallback.
list	implemented	Aggregate every scored account into a ranked CSV + per-account markdown dossier (gates, signals, contacts) for GTM hand-off.
personalize	implemented	Draft grounded outreach per contact from the detected signals — deterministic template with no key, rewritten by the model on the same evidence.

Quickstart (no keys)

Requires Python 3.10+. From the repo root:

# 1. Discover companies from a seed list (deterministic, no key) — or use
#    --brief "<ICP description>" once PERPLEXITY_API_KEY is set.
printf 'Acme Robotics, acme.example\n' > /tmp/seeds.csv
python3 skills/discover/scripts/discover.py --seeds /tmp/seeds.csv

# 2. Enrich a discovered (or supplied) account (local, no paid key)
python3 skills/enrich/scripts/enrich.py --slug acme-example --local

# 2b. Collect deep intent signals — scans the website, public ATS job boards
#     (Greenhouse/Lever/Ashby), and GitHub for the ICP's signal keywords
#     (e.g. hiring LangChain engineers). No key needed.
python3 skills/enrich/scripts/signals.py --slug acme-example

# 3. Classify is the LLM step — in Claude, run /gtm-icp:classify acme-robotics
#    Claude reads enrich.json + icp.criteria.json, grounds on corpus/, and
#    writes .gtm/acme-robotics/classify.json. Then score it:
python3 skills/classify/scripts/score.py --slug acme-robotics

# 4. Find contacts in a qualified account (no key -> persona targets; with
#    APOLLO_API_KEY -> real contacts mapped to the ICP's personas).
python3 skills/people/scripts/people.py --slug acme-robotics

# 5. Build the hand-off list across every scored account ->
#    .gtm/_report/accounts.csv + .gtm/_report/dossier.md
python3 skills/list/scripts/build_list.py

# 6. Draft grounded outreach per contact (no-key template; in Claude,
#    /gtm-icp:personalize acme-robotics then rewrites it on the same evidence).
python3 skills/personalize/scripts/personalize.py --slug acme-robotics

Per-account artifacts land in .gtm/<slug>/; the hand-off CSV + dossier land in .gtm/_report/ (all git-ignored). See skills/_shared/artifact-structure.md.

As a Claude plugin

Installed as a plugin, the stages are slash commands:

• /gtm-icp:discover <ICP brief | --seeds path> — find ICP-fit companies
• /gtm-icp:enrich <company | domain | account.json>
• /gtm-icp:classify <slug>
• /gtm-icp:people <slug> — find contacts in a qualified account
• /gtm-icp:list [--include-reject] — build the ranked CSV + dossier hand-off
• /gtm-icp:personalize <slug> — draft grounded outreach for an account's contacts
• /gtm-icp:pipeline <ICP brief | company | accounts.json> — runs discover→enrich→classify→score→people→list→personalize end to end