classify

Classify

Decide if an account fits the Ideal Customer Profile — and why — then score it deterministically. This is the stage where corpus grounding is the edge: the judgment is made against what has actually won, not a bare LLM guess.

When to use

• An account has been enriched (.gtm/<slug>/enrich.json exists) and needs an ICP fit verdict + score.
• The user wants to know not just the score but the evidence behind it.

Inputs

• .gtm/<slug>/enrich.json — the firmographic profile from enrich.
• .gtm/<slug>/signals.json — deep intent signals from enrich's signal scan (website / careers / GitHub keyword hits), if present. Each detected signal names the dimension it informs and carries cited evidence — use it.
• The ICP — read icp.criteria.json at the repo root (or the user-supplied path). It has four parts:
  • gates — boolean must-pass conditions. One failure → tier Reject.
  • dimensions — graded conditions, each worth up to max_points.
  • thresholds — tier_a / tier_b cutoffs on the normalized 0-100 score.
  • signals — keyword groups that map public-source signals to dimensions (consumed by enrich's signal scan; you read the results in signals.json).

Workflow

1. Ground the judgment. Retrieve relevant context for this account:

   • If k2_api_host / k2_api_key are configured, query the corpus via the bundled K2 stdio MCP shim for won-deal patterns, positioning, and case-study evidence matching the account's vertical/size.
   • Otherwise, read local corpus files under corpus/ (markdown notes on past wins, ICP rationale). Local grounding is the no-key path.
   • You are the agent here — there is no model call to orchestrate. Reason over the enrichment + the grounded evidence directly.

2. Evaluate the gates. For each gate, decide passed: true|false and write one line of evidence citing the enrichment field or grounded snippet. If a gate genuinely can't be evaluated from the evidence, mark it passed: false and say why — don't assume a pass. Any failed gate disqualifies the account.

3. Grade the dimensions. For each scoring dimension, award points_awarded between 0 and max_points (carry max_points through from the ICP), with one line of evidence. Partial credit is expected — reserve full points for strong, cited evidence. Do not invent firmographics absent from enrich.json. Use signals.json: for each dimension, pull the signal groups whose informs matches it. A found: true signal is concrete evidence to award points and should be quoted in the evidence line (e.g. "careers page hiring LangChain engineers → actively closing AI gap" for commercial_urgency). A found: false AI-product signal is itself evidence of a wide ai_gap. Prefer a cited signal over a firmographic guess.

4. Write the verdict to .gtm/<slug>/classify.json:

   {
     "company_name": "...",
     "gates": [
       {"key": "established",    "passed": true,  "evidence": "Founded 2014 (enrich.json)."},
       {"key": "not_ai_native",  "passed": true,  "evidence": "Core product is a TMS; AI is a recent add-on."}
     ],
     "dimensions": [
       {"key": "ai_gap",             "points_awarded": 24, "max_points": 30, "evidence": "..."},
       {"key": "data_workflow_moat", "points_awarded": 20, "max_points": 25, "evidence": "..."},
       {"key": "commercial_urgency", "points_awarded": 12, "max_points": 20, "evidence": "..."},
       {"key": "budget_access",      "points_awarded": 11, "max_points": 15, "evidence": "..."},
       {"key": "feasibility",        "points_awarded":  7, "max_points": 10, "evidence": "..."}
     ]
   }
   

5. Score deterministically:

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

   This reads classify.json and writes .gtm/<slug>/score.json with a normalized 0-100 score (sum of awarded points / sum of max points), a tier, and the breakdown. The score is pure arithmetic over your verdicts — no LLM judges the score, so it's reproducible and auditable. A failed gate forces tier Reject; otherwise the score maps to A / B / Nurture by the ICP thresholds (override with GTM_TIER_A / GTM_TIER_B).

6. Report the tier, score, any failed gates, and the dimensions that drove the score, with the evidence lines. The evidence is the point — a number with no grounding is what every commodity wrapper already produces.

Why grounding matters here

A bare LLM can guess ICP fit from firmographics. The defensible version cites your won deals and positioning — that's the difference between a replaceable wrapper and a tool that encodes what your team has learned. Lead with the evidence, not the score.

Offline scorer test: bash scripts/tests/test_score.sh.