Skills with receipts

Focused Agent Skills for specialized work — and benchmark evidence you can inspect.

What this is

Each skills/<name>/ directory is a self-contained Agent Skill: its SKILL.md is discovered first, and its references/assets are loaded only when relevant. Skills follow the Agent Skills standard; this repository's interaction protocol lives in AGENTS.md.

Quick start

git clone https://github.com/marshalfevzi/skills.git
cd skills
bun install
bun run validate
bun run new

bun run new starts the interactive documentation fetcher. Fetched source material lands in @tmp/<skill-name>/, and the resulting skill is created under skills/<name>/ with SKILL.md as its entrypoint.

Skills

1.0.0 is the version in each skill's SKILL.md; the repository package is separately versioned at 0.1.0. The Incus version is the version used by the committed fixture, not a compatibility ceiling. template is scaffolding and is not presented as a maintained skill.

Benchmarks

Eval case definitions live beside the skill at skills/incus-lxc/evals/evals.json; committed run evidence lives under benchmarks/incus-lxc/iteration-1.

This iteration runs three scenarios once with and once without the skill, so the pass rate does not distinguish the configurations; the observed signal is lower mean wall time and fewer avoidable validation errors in the with-skill runs. The recorded zero token values are not a quality metric.

Read the human report (benchmark.md), the machine-readable summary (benchmark.json), the test fixture (fixture.md), or any scenario directory: eval-deploy-web-with-profile, eval-dedicated-storage-pool, eval-isolated-project-network. Treat this as one iteration's snapshot, not a general performance guarantee.

bun-test-runner

Eval case definitions live beside the skill at skills/bun-test-runner/evals/evals.json; committed run evidence lives under benchmarks/bun-test-runner/iteration-1.

First-turn behavior evals for the Bun 1.4 update. Eval 4 (eval-ci-scaling-parallel) discriminates: with-skill runs pass all three Bun 1.4 assertions (--parallel, --shard/--timings, BUN_TEST_WORKER_ID); the pre-update snapshot fails all three. Evals 1–3 confirm the interrogation flow is unchanged. Read the human report (benchmark.md), the machine-readable summary (benchmark.json), the test fixture (fixture.md), or any scenario directory: eval-greenfield-time-sensitive, eval-jest-migration, eval-react-coverage. One run per configuration this iteration, so the spread reflects the eval mix, not repeatability.

Development

Add or update a skill

Run bun run new when external documentation should be fetched. Create a kebab-case skills/<name>/ from skills/template/SKILL.md; fill in name, description, version, and category. Keep long material in references/ or assets/, and make the frontmatter description sufficient for discovery. Run bun run validate before committing.

Add an eval

Keep case definitions beside the skill in skills/<name>/evals/, and commit each run under benchmarks/<name>/<iteration>/ with a human summary, a machine-readable summary, a fixture, feedback, and per-scenario with_skill/without_skill artifacts. When editing a skill, update its table row's content version, last-updated date, tested target, and evidence link. Do not put locally cloned repositories under benchmarks/; tmp/ is for ignored local clones.

Release

Run bun run release (standard-version); pushing a v* tag drives the draft GitHub Release workflow. Keep the repository package version distinct from each skill's content version.

Repository map

skills/incus-lxc/
├── SKILL.md
├── references/
└── evals/
benchmarks/incus-lxc/iteration-1/
├── benchmark.md
├── benchmark.json
├── fixture.md
├── eval-deploy-web-with-profile/
├── eval-dedicated-storage-pool/
└── eval-isolated-project-network/

benchmarks/bun-test-runner/iteration-1/
├── benchmark.md
├── benchmark.json
├── fixture.md
├── eval-greenfield-time-sensitive/
├── eval-jest-migration/
├── eval-react-coverage/
└── eval-ci-scaling-parallel/
scripts/
CONTRIBUTING.md
LICENSE

The root README is the index; per-iteration detail lives under benchmarks/.

Sources & thanks

• Agent Skills specification — the skill format this repository follows.
• Anthropic's skills examples — skill structure and phrasing reference.
• Incus documentation — source material for the incus-lxc skill.
• Linux Containers image server — fixture images used by the benchmark.
• awesome-readme — README design reference.

Contributing

See CONTRIBUTING.md for contribution guidelines and CODE_OF_CONDUCT.md for the community standards.

License

Licensed under the terms of the LICENSE file.

Last reviewed: 2026-08-26