enrich

Enrich

Turn a bare account record (a name + domain, or a discover result) into a firmographic profile the classify skill can judge against the ICP.

When to use

• The user hands you a company, a domain, or a list of accounts to qualify.
• A discover run produced raw accounts that need firmographic fill-in.
• You have an account but are missing size / industry / revenue to score it.

Inputs

A single account record as JSON with at least one of domain / website, plus any fields already known (company_name, industry, employee_count, …). The record lives at .gtm/<slug>/input.json (write it there first if the user gave you the account conversationally — see _shared/artifact-structure.md).

Workflow

1. Resolve the account slug (scripts/gtm_lib.py:slugify on the company name or domain) and make sure .gtm/<slug>/input.json exists.

2. Run the enrichment script:

   python3 ${CLAUDE_PLUGIN_ROOT}/skills/enrich/scripts/enrich.py --slug <slug>
   

   • With APOLLO_API_KEY set, this calls Apollo's organization-enrichment endpoint by domain and maps firmographics. Apollo is a synchronous REST API, so it works fine from a skill with no backend.
   • Without a key (or with --local), it runs the no-key fallback: firmographic-only, sourced from what the record already carries. This is the clone-and-run path — no paid vendor required to qualify on firmographics.

3. The script writes .gtm/<slug>/enrich.json and prints a one-line summary (slug, enrichment_source, artifact path). Surface the enrichment_source to the user so they know whether the data came from Apollo or the local path.

4. Collect deep intent signals (no paid key required):

   python3 ${CLAUDE_PLUGIN_ROOT}/skills/enrich/scripts/signals.py --slug <slug>
   

   This fetches the homepage, careers pages, public ATS job boards (Greenhouse / Lever / Ashby), and GitHub repos, then scans them for the keyword groups the ICP declares under signals — each group names the scoring dimension it informs. The ATS boards are public JSON APIs; the script first looks for a board link (boards.greenhouse.io/<slug>, jobs.lever.co/<slug>, jobs.ashbyhq.com/<slug>) embedded in the company's own careers HTML — that's the real slug, so it resolves boards a name-based guess can't (e.g. Sourcegraph's board is sourcegraph91). It falls back to guessing the slug from the name/domain when no link is in the HTML (JS-only careers pages won't expose one). Boards give structured per-posting text (where hiring signals actually live) and are surfaced in signals.json under hiring_boards (provider + slug + discovery provenance + posting titles/URLs) for the GTM rep. It writes .gtm/<slug>/signals.json with, per group, the matched_keywords, the evidence (source URL + snippet), and a found flag. Example: an ICP ai_hiring group (langchain, llm, …) that informs commercial_urgency will fire on a careers page hiring LangChain engineers — turning "they feel AI pressure" into a cited fact. Absence is evidence too: found: false on AI-product keywords legitimately widens an incumbent's ai_gap. GITHUB_TOKEN is optional (raises the GitHub rate limit); all fetches are best-effort and degrade to warnings.

5. For a batch, loop both scripts per account — one artifact dir each. (Entity fan-out across many accounts is the natural place to dispatch parallel subagents; keep each account's enrich isolated to its own slug.)

Notes

• Contacts vs firmographics. The local path qualifies on firmographics only. Verified contact data (direct dials, valid emails) requires a paid vendor key — that's a deliberate boundary, not a gap to paper over.
• Clay is async/webhook-based, so it does not fit a backend-free skill the way Apollo's sync REST does. Prefer Apollo (or a direct provider) here.
• Signals feed the score, not just the file. The point of signals.json is that classify cites it as per-dimension evidence — a hiring signal is what makes a commercial_urgency score defensible instead of a guess. Don't collect signals and ignore them.
• Both scripts are stdlib-only and offline-testable: bash scripts/tests/test_enrich.sh and bash scripts/tests/test_signals.sh.

This is the data stage. The judgment happens next in classify.