A curated collection of Claude skills. Load once, ship better output, save tokens.
This repo is a library of SKILL.md files that upgrade how Claude (and any Claude-powered tool) behaves — from decisions and writing voice to code review and skill management. Each skill auto-activates on matching triggers and ships with a documented first-response contract.
| Skill | What it does |
|---|---|
| claude-design | 7 design systems merged into one: animation physics, typography, OKLCH color, token architecture, Framer Motion, Bento 2.0, anti-slop rules + 20+ slash commands. |
| Skill | What it does |
|---|---|
| anti-hedge | Forced single-answer decisions. Bans "it depends", commits to one recommendation, explains what was ruled out and what would flip it. |
| real-problem | Root-cause question reframing. Catches XY problems and answers the question behind the question. |
| pre-mortem | Adversarial failure hunting. Stress-tests plans by building the strongest case against them before you commit. |
| constraint-solver | Treats constraints as design material. Finds elegant solutions when budget, time, tech, or regulation tightly restrict the space. |
| cross-domain-synthesis | Structural borrowing between fields — finds the mechanism in domain A that directly solves the problem in domain B. |
| iterative-refinement | 4–6 structured improvement cycles on any output without drifting from intent or regressing earlier gains. |
| gap-closer | Finishes 80%-done projects. Audits what exists vs. intended and outputs exactly ONE next action to unblock shipping. |
| Skill | What it does |
|---|---|
| voice-auth | Fingerprints your writing (sentence length, vocab, punctuation, filler) and rewrites output to sound like you — kills the AI smell. |
| tone-voice-mimicry | Copies any external voice — a brand, publication, or author — from sample text, stylistically indistinguishable. |
| structured-output | Enforces exact machine-readable output (JSON, YAML, CSV, markdown) with zero deviation, prose, or invented fields. |
| Skill | What it does |
|---|---|
| pr-review | Fans out 8 parallel review lenses (security, architecture, perf, dependencies, logging, testing, docs, a11y) into a severity-graded report. |
| red-teaming | Plays the attacker / adversarial user / skeptical engineer on code, systems, and plans. Surfaces the most dangerous findings first. |
| silent-failure-detection | Catches confident wrongness — surfaces hidden assumptions and probes calibration before hallucination causes damage. |
| constitutional-reasoning | Self-critique loop: evaluates its own output against principles, argues against itself, and revises until all principles hold. |
| Skill | What it does |
|---|---|
| large-doc-mastery | Precision synthesis of 50k–200k+ token contexts — codebases, PDFs, legal docs, transcripts — with perfect recall and cross-referencing. |
| claude-md-generator | Self-healing CLAUDE.md generator. Scans the codebase, writes a ≤50-line root CLAUDE.md, and patches gaps after every mistake. |
| Skill | What it does |
|---|---|
| context-router | Meta-skill that routes each query to only the relevant skills/MCP tools, then unloads heavy skills to keep context clean. |
| mcp-deduplicator | Detects overlapping MCP tools across servers, generates routing rules, and kills the "too many tools" tax. |
| trigger-auditor | Debugs skill triggers — fixes skills that won't fire, fire too eagerly, or conflict with each other. |
| context-injector | Ends the "let me explain again" tax — builds a structured context template and re-injects it per session. |
| precision-prompt-chaining | Chains dependent prompts (research → outline → draft → edit) with named variables, validation gates, and failure recovery. |
- Go to Claude.ai → Projects → Create new project
- In the project, go to Knowledge → Upload file
- Upload the
SKILL.mdfor the skill you want (e.g.skills/anti-hedge/SKILL.md) - Start a chat — the skill auto-activates on its triggers
Attach the SKILL.md to your first message in any Claude chat.
The SKILL.md files are fully portable:
- Cursor: paste the skill into
.cursor/rulesor Settings → Rules - Continue.dev: add under
customRulesin~/.continue/config.json - Claude API / LangChain / LlamaIndex: include the full text in your system prompt or as a retrieved document, prefixed with:
You are now using the <skill-name> skill. Follow every rule in the loaded skill file exactly.
skills:
- name: anti-hedge
path: skills/anti-hedge/SKILL.mdMany of these skills compose well together. A common setup:
- context-injector → persistent project context
- context-router → routes to the right skill per query
- claude-design → ships interfaces without slop
- pr-review + red-teaming → ships code safely
- voice-auth → output that sounds like you
Join the Discord to share skills, get help, and follow updates: discord.gg/MmsTNm8WF6
Star this repo if it saves you thousands of tokens and hours of prompt engineering.