Skip to content

README.md

Latest commit

 

History

History
122 lines (93 loc) · 6.29 KB

README.md

File metadata and controls

122 lines (93 loc) · 6.29 KB

skills-from-the-artificer

A collection of Claude Code skills, distributed as drop-in directories. Each top-level folder is a self-contained skill — install it with the skills CLI, or copy it into ~/.claude/skills/ and restart Claude Code.

Skills

Skill What it does
skills-from-the-artificer/ Meta dispatcher. Given a change, fix, diff, design, or decision, classifies it and points you at the relevant law/principle skill(s) from the 24-law collection below. The entry point for "which laws apply here?" and the Bug Remediation Step 2 cross-reference.
cost-tier-routing/ Routes work to the cheapest model that can do it correctly — haiku for search/IO, sonnet for coding, opus for orchestration. Ships a skill and three model-pinned subagents.
cost-routing/ Top-level dispatcher that classifies requests into scout / coder / architect tiers and ships dispatch templates + a PreToolUse warning hook. Companion form to cost-tier-routing.
rubber-duck/ Interactive rubber-duck debugging session. Forces you to reconstruct your mental model of the bug from scratch, which is where the bug usually surfaces.
test-first-bugfix/ Test-driven bug fixing — reproduce the bug as a failing test before you touch the fix. Catches "fixes" that don't actually fix anything and prevents regressions.
trust-but-verify/ Re-validates every claim a subagent hands back against a primary source — the code, docs/ADRs, the memory dir, context7, or a language spec — before you act on it. Treats a report as a lead, not a fact: nothing is verified until a source was opened and quoted.
ci-preflight/ Pre-push checklist that prevents CI whiplash — multiple red pushes that could have been caught locally. Guards against three anti-patterns: missing registration surfaces when adding new shipped files, skipping npm test before pushing, and guessing at cross-platform fixes instead of diagnosing the root cause.

Laws of software (24 reference skills)

A pack of software-engineering laws, each loaded as an on-demand reference skill. Every law is its own top-level directory:

choose-boring-technology/ · conways-law/ · cunninghams-law/ · doerrs-law/ · fitts-law/ · galls-law/ · goodharts-law/ · greenspuns-tenth-rule/ · hofstadters-law/ · hyrums-law/ · kerckhoffs-principle/ · kernighans-law/ · knuths-optimization-principle/ · lady-lovelaces-objection/ · leaky-abstractions/ · linuss-law/ · moores-law/ · norvigs-law/ · parkinsons-law/ · peter-principle/ · postels-law/ · shirky-principle/ · wirths-law/ · zawinskis-law/

Install (recommended: the skills CLI)

The skills CLI discovers each top-level skill and installs it for you. Add -g to install globally (user-level, into ~/.claude/skills/) instead of into the current project.

# install everything
npx skills add The-Artificer-of-Ciphers-LLC/skills-from-the-artificer --all -g

# or choose interactively
npx skills add The-Artificer-of-Ciphers-LLC/skills-from-the-artificer -g

# or install specific skills
npx skills add The-Artificer-of-Ciphers-LLC/skills-from-the-artificer -g --skill conways-law,rubber-duck

Install (manual)

Each skill is also a cp -r away:

# pick the ones you want
cp -r rubber-duck ~/.claude/skills/
cp -r test-first-bugfix ~/.claude/skills/
cp -r conways-law ~/.claude/skills/

# or grab every law at once
cp -r *-law *-principle *-rule *-objection *-technology *-abstractions ~/.claude/skills/

Restart Claude Code. The skill discovery scan runs at session start.

Install (skills that ship agents too)

cost-tier-routing/ is unusual — it includes companion subagents that have to land in ~/.claude/agents/, not ~/.claude/skills/. Its README.md documents the two-step install. The skill alone is inert without the agents because the model: pin lives in the agent files.

cd cost-tier-routing
mkdir -p ~/.claude/skills/cost-tier-routing ~/.claude/agents
cp SKILL.md ~/.claude/skills/cost-tier-routing/
cp agents/*.md ~/.claude/agents/

How skills work

A skill is a directory containing a SKILL.md with YAML frontmatter — name and description — followed by the prose Claude Code reads when the skill triggers. The description is the trigger surface: Claude scans loaded skill descriptions against the current task and invokes a skill via the Skill tool when the description matches. Skills can ship supporting files (eval suites, reference scripts, sub-skills) alongside SKILL.md.

See the Claude Code skills docs for the canonical reference.

Evals

Some skills include an evals/evals.json file with prompts and expected behaviors. These are consumed by the skill-creator workflow to benchmark whether a skill triggers correctly and produces the expected routing decisions. The format is intentionally minimal:

{
  "skill_name": "<name>",
  "evals": [
    {
      "id": 0,
      "prompt": "<user prompt that should trigger the skill>",
      "expected_output": "<what Claude should do, prose>",
      "files": []
    }
  ]
}

Contributing a skill

  1. Create a top-level directory named after the skill (kebab-case).
  2. Drop in a SKILL.md with the frontmatter name: and description: fields, followed by the body Claude should read.
  3. Optionally add evals/evals.json with trigger prompts and expected behaviors.
  4. Optionally add a per-skill README.md documenting anything unusual about the install (e.g. companion agents).
  5. PR.

License

Skills in this repo are MIT-licensed unless a skill's own directory says otherwise.