New skill: evidence-based-reviews (content)

[adunn08]

A generic, public method for honest product reviews and buying guides without fabricated first-hand experience. Content category, display_order 13.

The skill

• The anchor rule: never claim hands-on time that did not happen. Everything else makes honest content strong enough that the lie is unnecessary.
• Four evidence tiers with the work each requires: verified manufacturer specs (cited to the maker's page with the verification step, not aggregator copies), owner-experience synthesis at scale (recurrence over volume, sample-size candor, splits as findings, no cherry-picking), expert triangulation (named sources, disagreements surfaced not averaged), and hands-on only when true with the upgrade-and-mark pattern (never backfilled).
• The per-piece methodology block: placed with the criteria, written for readers, with the mandatory hands-on line even when the answer is none claimed.
• Structure discipline: criteria before picks, a named drawback per pick, dated what-changed updates.
• Google reviews-system alignment: what demonstrable original analysis concretely means (the synthesis, comparative tables from verified data, decision frameworks; not paraphrased spec sheets).
• FTC alignment: plain-language disclosure traveling with the recommendation, never footer-buried.
• Schema judgment: markup is a claim; ItemList and Article for synthesis guides, Review only when tier 4 is true, per piece.
• Two reference files: the tiers in depth, and the fillable methodology block with a worked example and placement rules.

Catalog discipline

• Structure matches the exemplars (content-and-copy, seo-onpage): same frontmatter shape, section order, when-not-to-use convention. Every when-not-to-use name resolves to a real skill (content-and-copy, brand-voice, seo-onpage, content-brief-authoring).
• Description: 910 characters, under the 1024 platform-port cap.
• README catalog regenerated via scripts/generate_readme_catalog.py --write (103 skills, 448 reference files; all counts derived, never hardcoded); the --check sync gate passes.
• SKILLS.lock regenerated via tools/gen_skills_lock.py; the lock diff is exactly the new skill's five-line entry and the local --check passes. CORRECTION: the first version of this PR body flagged gen_skills_lock.py and SKILLS.lock as missing from the repo; that was wrong (the check ran against a stale local main checkout before branching). A Windows hashing nuance surfaced while fixing it: the tool hashes raw bytes, so the lock must be generated from LF working files; this machine's checkout now uses repo-local core.autocrlf=false to keep regeneration CI-parity.
• Em-dash grep across the skill folder: zero. Generic method throughout; no Bowhunt America, RampStack-brand, or vertical specifics.
• Diff scoped to skills/evidence-based-reviews/, SKILLS.lock, and the generated README sections.

Held for Andy's squash-merge.

🤖 Generated with Claude Code