Refactor slash commands into a Claude Code skill architecture

[larnsce]

Background

Commands and skills have converged in Claude Code; skills additionally support supporting files with progressive disclosure, frontmatter control (disable-model-invocation, allowed-tools), and cleaner packaging. Research (July 2026) found no equivalent existing implementation to borrow wholesale; the architecture is ours to design. A premortem on this refactor identified the main risks; see premortem-report-20260708-053608.html and premortem-transcript-20260708-053608.md.

Adoption/distribution by other reviewers is explicitly out of scope for now; a marketplace can come later if ever needed.

Approach

Prototype first, then convert:

1. Convert the read-only /review-status command into a skill as a prototype. Verify firsthand rather than trusting secondhand docs: frontmatter field behavior, whether reference files resolve reliably when the skill runs from a target package directory, and context loading behavior.
2. If the prototype holds, convert the remaining commands one skill per entry point (review-package, review-issue, create-next-issue, review-complete, create-release), each with disable-model-invocation: true so nothing auto-triggers.
3. Move long reference material (canonical checklists from Consolidate review checklists into a single source of truth #5, templates, standard _pkgdown.yml) into skill supporting files.

Kill criterion

If the prototype cannot reliably load its reference files from a target package directory after one day of testing, stop the skill migration and keep the command files, deduplicated per #5, instead.

Tasks

• Prototype: review-status as a skill, tested from inside a real data package repo
• Document verified platform behavior: what frontmatter fields actually do in the installed Claude Code version
• Convert remaining commands to skills
• Retire commands/: replace files with deprecation stubs pointing at the skills
• Update README install instructions

Depends on #4 and #5: do not migrate contradictory content.

[larnsce]

Implemented on dev in cb2d9d3. All six entry points are skills (review-package, review-issue, create-next-issue, review-status, review-complete, create-release), each with disable-model-invocation: true; review-pr folded into review-issue step 7. Shared references live in skills/pkgreview-core/, loaded via CLAUDE_SKILL_DIR paths. commands/ files are deprecation stubs; README install switched to copying skills/ (symlinks execute but skill discovery has known bugs with them, see anthropics/claude-code#25367 and #14836).

Verified platform behavior (v2.1.170 docs): disable-model-invocation restricts to user slash invocation; ARGUMENTS substitution works in skill bodies; CLAUDE_SKILL_DIR resolves regardless of cwd; supporting files load on demand; skill content stays pinned in context once invoked (relevant to #8).

Remaining before the kill criterion is retired: test /review-status (the prototype) from inside a real data package repo, then run the fixture gate (#9).

[larnsce]

Prototype test passed (2026-07-08): user ran /review-status from a real data package repo after installing the skills by copy; reference files loaded correctly via CLAUDE_SKILL_DIR from a foreign working directory. Kill criterion retired; the skill migration stands. All tasks checked off. Remaining gate before merging PR #14 is the fixture review (#9).

[larnsce]

Delivered: implemented on dev via PR #14 (merged 2026-07-08), validated by the first production review (openwashdata/fslogisticskampala v1.0.0), and released as review standard v1.0.0 via PR #26.