harness-skill-compare

Use for a substantive comparison of skill triggers, behavior or curation choices. Scale narrow overlap questions to the relevant branches.

SKILL.md

GitHub

Harness Skill Compare

Compare selected skills at the depth the question requires. For a narrow overlap question, read descriptions, relevant body branches and their direct dependencies. For a full behavioral comparison, inventory all shipped instructions, metadata and scripts, distinguishing inspected code from executed behavior.

Consult harness_lifecycle/registry/ for classification and prior decisions: INDEX.md for coordinates, skill-records/ for our skills' lineage, history and earlier reviews, references/<harness>.md for rulings on reference capabilities. Use configured bounded delegation only when helpful; no mandatory subagent-choice question applies.

For a durable comparison that decides a merge, upgrade or rebuild of one of our skills, write a skill review: harness_lifecycle/registry/reviews/<yyyy-mm-dd>-skill-<our-skill>.md (registry/README.md §5a). Its parts, in order:

  1. Set: our skill and each upstream counterpart, with repo, path and the pinned commit read.
  2. Components: each behavioral component (step, gate, template, rule, heuristic, red-flag table) with file:line. A heading is not a component.
  3. Matrix: one row per component, one column per skill: ✓ present, ~ variant, — absent.
  4. Differences: for every ~ row and differing ✓ row, how each skill realises it and which mechanism is stronger, with the reason.
  5. Verdict: substitutes or complements; what to take, from where, and what to leave, each with its reason.

A narrow answer stays in chat and creates no file. State coverage and unverified behavior; description text alone does not prove actual invocation frequency.

Comparison is read-only on the compared skills. Rulings belong to authorized harness-evaluate work, which cites the review in each event's evidence.