Growth Skills

Installable, testable Agent Skills for growth engineering.

Most “growth prompts” jump from a vague idea to content or a dashboard. This repository focuses on the contracts in between: observable surfaces, source evidence, campaign identity, event continuity, and explicit measurement boundaries.

Skills

Skill	Job	Deterministic output
utm-audit	Make campaign links reviewable without inventing missing meaning	Per-link pass/warn/fail, safe normalization, PII redaction
seo-surface-audit	Audit what deployed pages actually expose to crawlers	Indexability conflicts, sitemap/canonical findings, batch duplicates
trend-radar	Refresh and rank source-backed social signals	draft_now / research / review / watch / reject lanes
measurement-contract-audit	Check whether a funnel can be measured before querying it	ready / partial / blocked, event coverage, identity and attribution gaps

Together they form a small evidence chain:

SEO or social observation
  → source-backed opportunity
  → content / campaign ID
  → UTM contract
  → event + identity + metric contract
  → later production readback

No Skill in this repository pretends the final arrow happened.

Try it

Requirements: Node.js 20 or newer. Python 3.9 or newer is required only for deterministic package builds.

npm test
npm run audit:utm-example
npm run audit:seo-example
npm run audit:trend-example
npm run audit:measurement-example
npm run pack
npm run verify:packages

The checkers read JSON, make no network requests, and do not mutate external systems. Some intentionally exit with status 2 when the supplied artifact is blocked; they still print the complete report.

Design rules

• Evidence before advice. Runtime observations, source URLs, timestamps, event contracts, and explicit unknowns are inputs—not details to fabricate.
• Deterministic core. Scoring, normalization, coverage, and severity live in scripts that can be tested and reproduced.
• Facts stay separate from inference. A high trend score is not a forecast; a clean SEO snapshot is not proof of indexing; tidy UTMs are not attribution.
• Privacy survives reporting. Obvious contact data is redacted rather than copied into a cleaner-looking artifact.
• Mutation stays elsewhere. These Skills audit and prepare decisions. They do not publish posts, edit pages, replace links, create events, or write production data.

Repository shape

skills/<name>/
  SKILL.md
  scripts/
  evals/
  examples/
  references/        # when a Skill needs a maintained taxonomy
tests/
scripts/pack_skills.py
benchmarks/

Every Skill remains independently understandable and installable. npm run pack writes one reproducible .skill archive per Skill plus manifest.json and SHA256SUMS; npm run verify:packages compares every archive member with the checked-in source.

Evidence

• The test suite covers deterministic ranking and reports, identity continuity, sitemap/indexability contradictions, UTM normalization, privacy redaction, input immutability, Skill contracts, relative links, and reproducible package bytes.
• utm-audit has an initial one-run-per-prompt directional evaluation: 12/12 assertions with the Skill and 9/12 without it. See benchmarks/utm-audit-v0.1.md.
• trend-radar has a one-run directional evaluation: 6/6 assertions with the Skill and 3/6 without it. See benchmarks/trend-radar-v0.2.md.
• The remaining Skill eval prompts are checked in. Broader repeated behavior benchmarks are still required before making comparative model-quality claims.

Status

0.2.0 is a working foundation, not a claim that growth can be automated from four JSON files. The next useful additions are a GSC opportunity audit and a cross-Skill content/campaign ID contract, followed by production readback adapters that remain read-only by default.