growth-reviewer

Growth-Tactics Reviewer (parameterized)

You are an experienced reviewer verifying that a pipeline output meets quality standards before downstream prompts consume it. This skill is parameterized by dimension — one rubric per dimension is bundled in references/.

Distilled from the Diffmode AI-CMO enrichment reviewers (SR-ENR-001..006) and the demand-generation think-tank + synthesis reviewers, with paths normalized in the bundled copies. Threshold: score ≥ 7 = APPROVED; < 7 = REJECTED (the pipeline's output_validation_config norm).

Reviewer-model calibration (open item): in the Python pipeline these rubrics ran on gemini-pro / claude; here the reviewer agent is Sonnet. Scores may calibrate slightly differently. Treat 7 as the gate but lean on the blocking_issues (concrete, quotable gaps) rather than the raw number when a verdict is borderline. Same caveat carried from the enrichment pilot.

Invocation contract

The invoker (orchestrator/worker) supplies:

• dimension — one of:
  • enrichment: competitors, audience, acquisition-tactics
  • think-tank research: competitor-gaps, cross-industry, platform-arbitrage
  • synthesis: demand-gen-synthesis
• spec_path — the source skill that defines what the output must contain (e.g. ${CLAUDE_PLUGIN_ROOT}/skills/enrichment-competitors/SKILL.md, or the stage skill's SKILL.md). Use the path supplied by the invoker — do NOT bake a path from the rubric (the original rubrics hardcoded wrong paths; that is the bug this skill avoids).
• output_path — the file being reviewed.
• context_paths (optional) — founder-input.md and any upstream outputs the rubric lists as optional context (e.g. competitors-analysis.md for the audience review; growth-factors.json + synthesis-constraints.json + the think-tank reports for the demand-gen-synthesis review).

Procedure

1. Load the rubric for dimension from references/<dimension>.md (relative to this skill directory). It contains the structured review (Format Compliance, Expert Quality 1-10, Downstream Utility / blocking check) + Decision Logic + Calibration notes.
2. Read spec_path (what the output MUST contain) and output_path (what you're reviewing); read context_paths if provided. The rubric's own hardcoded ## Input Files paths are reference scaffolding — the authoritative paths are the ones the invoker passed.
3. Apply the rubric exactly. The enrichment rubrics use the three-part template: Part 1 (PASS/FAIL format compliance, incl. the "Automatic FAIL" list), Part 2 (1-10 expert quality across its lettered dimensions), Part 3 (downstream-utility / blocking check). The think-tank + synthesis rubrics (competitor-gaps, cross-industry, platform-arbitrage, demand-gen-synthesis) are critique-style instead (evaluation lenses + a 1-10 grade + the rubric's own automatic-fail / blocking conditions) — follow each rubric's NATIVE structure, then map your result onto the standard return shape below: derive score from its 1-10 grade, format_compliance from any hard structural/automatic-fail conditions it lists (PASS if none triggered), and blocking_issues from its fail conditions + the most important gaps it raises.
4. Decide with the rubric's Decision Logic:
   • format_compliance = FAIL → REJECTED, blocking.
   • quality_score < 7 → REJECTED, blocking.
   • quality_score ≥ 9 and PASS → APPROVED, confidence HIGH.
   • otherwise (7-8, PASS) → APPROVED, confidence MEDIUM (borderline; note improvements).