minimal-skills

Claude Code skills for the Minimal developer system, packaged as a plugin, plus the eval pipeline that keeps them honest. Each skill teaches Claude Code how to drive one slice of Minimal (setup, configuration, networking, diagnostics), and every skill is continuously tested for trigger accuracy, answer quality, and drift against the published docs.

claude plugin marketplace add gominimal/minimal-skills
claude plugin install minimal@gominimal

Or run /plugin inside the Claude Code REPL and install minimal from the gominimal marketplace.

Two tiers, two canaries. The binding contract for cases, checks, and the runner CLI is evals/SCHEMA.md; the methodology follows Testing Agent Skills.

• PR text tier: every pull request runs a no-secrets URL lint, then the text-tier suite (does the right skill trigger, does no skill trigger on negative cases, does the response pass its checks). One trial per case; regression cases gate the merge.
• Nightly functional tier: the full suite (3 trials) on runners with a real Minimal installed from both the stable and nightly channels, so a Minimal release that breaks a skill surfaces within a day.
• Obsolescence canary (nightly, advisory): the text tier re-run with the skills not installed. When the bare model passes everywhere, a skill has stopped paying rent and should be slimmed or retired.
• Docs-drift canary (nightly): the published minimal.dev pages the skills link are hashed and compared to evals/docs-snapshots.json; a changed page fails the run so the skill gets re-checked against it.

Failures of blocking nightly jobs open or update a single tracking issue.

Needs the claude CLI on PATH, authenticated (locally via claude login; CI passes a CLAUDE_CODE_OAUTH_TOKEN secret generated with claude setup-token) (npm install -g @anthropic-ai/claude-code); lint-urls needs neither.

just lint-urls        # liveness-check every docs URL in the skills
just eval             # text tier, all skills, 1 trial
just eval-skill minimal-setup   # text tier, one skill
just triggers         # fast trigger-accuracy pass
just functional       # functional tier; needs a local Minimal install
just obsolescence     # text tier with the skills not installed
just report           # open the latest JSON report

.claude-plugin/       plugin and marketplace manifests
skills/               one directory per skill, each with a SKILL.md
evals/                eval harness: runner.py, checks.py, cases/, SCHEMA.md
.github/workflows/    pr-evals.yml (PR gate), nightly-evals.yml (nightly tier)
justfile              local dev-loop recipes

• Skills link the published docs at minimal.dev instead of restating them; the URL contract lives in evals/SCHEMA.md.
• minimal-networking is experimental: its surface has no public docs (that is intentional), so it carries inline knowledge and may change or break without notice.

MIT, see LICENSE.

Note: the optional LLM style judge (--judge) uses the Anthropic SDK and needs ANTHROPIC_API_KEY; the default pipeline authenticates the claude CLI alone and never requires it.