TOOL-002 — Dev-Persona ICP Enricher

Tier 3 Specialist Tool · Stateless · Augments AGT-201 ICP scoring with developer-persona signals · Closes the dev-buyer side of the Domain 3 gap

Tier 3 · Tool Specced · v29 Domain 3 · API/Dev GTM Haiku

Purpose

AGT-201's 6-dimension ICP scorer captures firmographic, vertical, revenue, tech-stack, growth, and intent signals — but materially under-weights developer-buyer signal. For API-first products with developer-led adoption, the buying signal is technical: who's hiring engineers with specific stacks, whose GitHub orgs are public-active, who's experimenting with the product via API keys before commercial conversation. TOOL-002 takes those signals and produces a structured dev-persona enrichment that AGT-201 consumes alongside its existing dimensions.

Closes the dev-persona side of the Domain 3 gap. AGT-201 keeps its existing 6-dimension model unchanged; the dev-persona enrichment is a parallel signal stream, not a replacement.

Input schema

{
  "account_id": "uuid",
  "account_basics": {
    "company_name": "string",
    "domain": "string",
    "employee_count": 0,
    "industry": "string"
  },
  "technical_signals": {
    "job_postings": [                      // optional — observed in last 90 days
      { "title": "string", "stack_keywords": ["string"], "posted_date": "ISO 8601" }
    ],
    "github_org_activity": {                // optional — public-only data
      "org_name": "string",
      "public_repo_count": 0,
      "trending_languages": ["string"],
      "recent_topic_keywords": ["string"]
    },
    "technographic_data": {                 // from AGT-204 enrichment if available
      "detected_stack": ["string"],
      "infrastructure_signals": ["string"]  // e.g., "uses Stripe", "deploys via GitHub Actions"
    },
    "product_telemetry": {                  // optional — UsageMeteringLog or upstream
      "api_keys_active": 0,
      "first_api_call_at": "ISO 8601",
      "endpoint_diversity_score": 0.0      // 0-1, how many distinct endpoints touched
    }
  },
  "current_icp_score": {                   // from AGT-201, for context
    "tier": "T1" | "T2" | "T3",
    "dimension_breakdown": {
      "firmographic": 0, "vertical": 0, "revenue": 0,
      "tech_stack": 0, "growth": 0, "intent": 0
    }
  }
}

All technical_signals subfields are optional — the tool degrades gracefully when data is sparse, and the output reflects which signals were available vs. missing.

Output schema

{
  "tool_call_id": "uuid",
  "account_id": "uuid",
  "dev_persona_score": 0,                  // 0-100 composite
  "dev_persona_tier": "high" | "medium" | "low" | "none",
  "buyer_persona_classification": {
    "primary": "developer" | "platform_team" | "compliance_engineer"
             | "data_engineer" | "non_technical_buyer" | "unclear",
    "confidence": "high" | "medium" | "low"
  },
  "signals_present": {                      // which signals contributed
    "job_postings_signal": "strong" | "weak" | "absent",
    "github_signal": "strong" | "weak" | "absent",
    "technographic_signal": "strong" | "weak" | "absent",
    "telemetry_signal": "strong" | "weak" | "absent"
  },
  "key_drivers": [                          // top 3 drivers of the score, with citations
    { "driver": "string", "evidence": "string", "input_field": "string" }
  ],
  "agt_201_recommended_adjustment": {       // suggested dimension adjustments for AGT-201
    "tech_stack_dimension_adjust": 0,      // -15 to +15; bounded
    "intent_dimension_adjust": 0,          // -15 to +15; bounded
    "rationale": "string"
  },
  "ungrounded_assumptions": ["string"],     // explicit list, parallels TOOL-001
  "data_completeness": "high" | "partial" | "sparse",
  "tool_metadata": {
    "model": "claude-haiku-4-5",
    "input_tokens": 0, "output_tokens": 0,
    "cost_usd_estimate": 0.0,
    "latency_ms": 0
  }
}

Bounded adjustment magnitude. The tool can recommend AGT-201 dimension adjustments up to ±15 points on tech_stack and intent dimensions only. It cannot recommend changes to firmographic, vertical, revenue, or growth dimensions — those are not under TOOL-002's purview. AGT-201 retains the right to ignore the adjustment recommendation.

Called by

Caller	Invocation context
AGT-201 ICP Scorer	On account update events that include technical signals (job posting added, GitHub activity refreshed via AGT-204, product telemetry signal change). AGT-201 calls TOOL-002, ingests the recommended adjustment as one input alongside its 6-dimension model, decides whether to apply.
AGT-901 Pipeline Brain	Cross-account analysis of dev-persona patterns. "Are we under-tiered in API-first companies hiring backend engineers?" Brain calls TOOL-002 across a target list to identify dev-persona-rich accounts that AGT-201 ranked low.
AGT-204 Lead Enrichment	As an enrichment subroutine. AGT-204's existing 6-category enrichment may invoke TOOL-002 when technical signals are present, before the rescore trigger fires AGT-201.

Design principles

Augment, don't replace. AGT-201's 6-dimension model stays the canonical ICP signal. TOOL-002 is a parallel enrichment that AGT-201 chooses to integrate — the existing scoring logic is unchanged.
Sparse-data friendly. Technical signals are uneven by company. Most accounts won't have GitHub org data; many won't have product telemetry. The tool degrades gracefully and reports data_completeness honestly.
Bounded magnitude. Dev-persona signal is real but not infinite — capped at ±15 points on two dimensions. A high-grade developer fingerprint shouldn't single-handedly flip a Tier 3 to Tier 1; the firmographic and revenue dimensions still matter.
Distinguish persona from intent. A company can have strong developer presence without buying intent for our specific product. The output separates buyer_persona_classification (who they are) from intent_dimension_adjust (whether they're showing buying signal).
Honest unclear classification. When signals point in conflicting directions, the tool returns "unclear" rather than fabricating a confident classification.

Cost ceiling

Constraint	Value
Per-call input budget	10K tokens (account + signals + current ICP score; bounded narrow context)
Per-call output budget	1K tokens (structured enrichment output)
Default model	Haiku — classification task with bounded scope; Sonnet tested with no measurable quality lift
Per-call cost estimate	~$0.01–$0.02 per call at Haiku pricing
Monthly cap (default)	$200/mo — bounds usage to ~10,000 calls/month
Frequency expectation	High — called on AGT-204 enrichment events, AGT-201 rescore events, AGT-901 batch analyses. Most monthly usage will come from AGT-204 routine enrichment.

Eval criteria

Criterion	Measurement	Pass threshold
Schema compliance	Output validates against output schema	100% (hard)
Persona classification accuracy	15-account retrospective set with known buyer personas (from won deals); % where primary classification matches	≥ 70%
Sparse-data graceful degradation	Tool produces useful output when only 1-2 of 4 signal categories present (vs. fabricating confidence)	`data_completeness` honestly reflects sparse state; confidence is `low` or `medium` not `high`
Adjustment magnitude calibration	Recommended adjustments correlate with retrospective dev-persona quality (won deals at dev-persona accounts)	Pearson correlation ≥ 0.4 across the eval set
Hallucinated signals	% of `key_drivers` citing input fields that don't exist or aren't populated	0% (hard)
P95 latency	End-to-end tool call	≤ 2s — tool is called in line with AGT-204 enrichment, latency budget tight

Failure modes

Symptom	Cause	Action
Tool inflates dev-persona scores when signals are sparse	Confidence calibration drift	Eval catches. Tighten "honest unclear classification" instruction in prompt; add few-shot examples of sparse-data outputs that correctly returned low confidence.
Tool flips firmographically-low accounts to T1 via adjustment	Magnitude bound violated, or AGT-201 over-applying recommendation	The ±15 cap is enforced in the output schema. If the issue is AGT-201 stacking adjustments, that's an AGT-201 calibration concern, not TOOL-002.
High latency in batch invocations from AGT-901	Tool called per-account synchronously in a 100-account batch	Caller responsibility — AGT-901 should batch-parallelize calls. Tool itself is stateless and fast per-call.
Persona classifications drift over time	Tech stack signals aging (yesterday's "developer" stack is today's mainstream)	Quarterly catalog review; refresh prompt examples with current dev-persona signals.
Cost spike	AGT-204 enrichment running on too-frequent cadence	Cost cap blocks runaway. Audit AGT-204 invocation cadence; per the existing v22 spec it should be T1 daily, T2 weekly, T3 monthly — if drifted, reset.

Interaction with AGT-201 ICP scoring

AGT-201's spec is unchanged. The integration point is: AGT-201 reads tech_stack_dimension_adjust and intent_dimension_adjust from the most recent TOOL-002 output for the account (logged in EnrichmentLog with the AGT-204 enrichment record), and treats them as additive deltas to its computed dimension scores. AGT-201 retains discretion to ignore the adjustment if other signals contradict.

For backwards compatibility, accounts with no TOOL-002 enrichment are scored exactly as before (deltas default to 0). The dev-persona enrichment is purely additive — no existing T1/T2/T3 tier is downgraded by its absence.