claude-skills

License: MIT Claude Code claude.ai

A collection of skills for Claude Code, claude.ai, and other Claude agent surfaces.

Skills

Skill What it does
token-efficiency Applies to every response, not one task type — code, writing, research, analysis, and plain chat alike, on any surface. Governs how Claude reasons, writes, uses tools, and manages its own context so output and context growth stay small without losing accuracy — trimming filler, redundant restatement, unnecessary verification loops, redundant re-reads, and bloated code.

token-efficiency

The core idea: the goal is never "say less" — it is the smallest set of high-signal tokens that gets the outcome right, measured across the whole session, not one message. Cutting a needed clarifying question or a verification run to save tokens is a false economy the skill explicitly forbids.

What you get

  • The skill body (SKILL.md) — seven sections of standing habits: efficient reasoning, prose, code, agentic tool use, context-window management, guardrails on where not to cut, and a pre-send self-check.
  • An always-on hook layer (hooks/), Claude Code only — re-injects the core contract every turn, flags concrete tool-use waste (redundant re-reads, needless re-runs) as it happens, and prints a per-task usage report. Detects explicit requests for depth in 12+ languages and backs off the brevity rules when it sees one.
  • Worked before/after examples (references/examples.md) — code and non-code: repo edits, contract review, plan comparison, email drafting.
  • An A/B eval harness (evals/) for measuring the skill against a no-skill baseline on your own machine.

Measured results

Four eval tasks (two agentic repo tasks, two everyday non-developer tasks), each run with and without the skill. All runs pass 100% of the content assertions — the skill does not trade correctness for brevity. The difference is size:

Task Without skill With skill Reduction
Explain config wiring (explicit depth requested) 644 words 421 words −35%
"Which internet plan should I get?" 255 words 122 words −52%
"Anything to worry about in these SaaS terms?" 441 words 302 words −32%
Targeted repo edit + test run 61 words 50 words −18%

Same coverage of the substance in every case — the guardrails kept the safety caveats and the explicitly-requested depth intact. Reproduce with the harness in evals/.

Installing

Claude Code

git clone https://github.com/denfry/claude-skills.git
cp -r claude-skills/skills/token-efficiency ~/.claude/skills/token-efficiency

Claude Code picks up skills from ~/.claude/skills/ (user-level) or .claude/skills/ (project-level) automatically — no restart required.

Then, optionally, install the always-on hook layer:

python ~/.claude/skills/token-efficiency/hooks/install.py --dry-run   # inspect first
python ~/.claude/skills/token-efficiency/hooks/install.py

The install is additive and idempotent — existing hooks on the same events are kept, re-running replaces this skill's entries instead of stacking duplicates, and --uninstall removes them cleanly. See hooks/README.md for what each hook does and the env vars that tune them.

claude.ai / desktop / mobile

Zip the skill folder, rename the archive to token-efficiency.skill, and upload it under Settings → Capabilities → Skills. Everything except the hooks/ layer works there — hooks require the Claude Code harness, but the skill body does not depend on them.

Structure

skills/
└── <skill-name>/
    ├── SKILL.md          — required: YAML frontmatter (name, description) + instructions
    ├── CORE.md           — optional: the always-on subset, injected by a hook each turn
    ├── hooks/            — optional: harness-level enforcement + install.py
    ├── evals/            — optional: eval prompts and a measurement harness
    └── references/       — optional: supporting docs loaded on demand, not on every trigger

SKILL.md frontmatter is what Claude uses to decide when to consult a skill, so the description field is written to trigger reliably rather than just to summarize the skill. The body stays under ~500 lines; anything longer (worked examples, per-domain references) lives under references/ and is linked from the body instead of inlined.

See Anthropic's skill authoring guide for the full format.

Contributing

New skills or improvements to existing ones are welcome via PR. Keep the same structure: a lean SKILL.md with a pushy, specific description, and any bulky reference material split out under references/.

License

MIT