ARC Skills
a-canary/arc-skills
Skills curadas e sem dependências para automação em Claude Code
Ver no GitHubColeção curada de skills sem dependências externas para Claude Code e harnesses similares. Cada skill tem propósito claro e valor distinto, organizadas em categorias como guardrails comportamentais, higiene de código, workflows de engenharia, e integrações de domínio específico.
Instalação
npx convez-ai a-canary/arc-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automatizar workflows de engenharia (TDD, PRDs, triagem)
- Aplicar guardrails comportamentais e higiene de código
- Gerenciar infraestrutura e integrações com sistemas externos
Veredito

Repositório robusto com skills bem estruturadas e documentadas, zero-dependency por design. Adequado para quem quer estender Claude Code com comportamentos específicos sem overhead de dependências.
README
A tradução para o português deste README ainda está na fila.
arc-skills
Zero-dependency, curated skills for Claude Code and similar harnesses.
Every skill here earns its slot: clear purpose, distinct value, no overlap. Lifted skills include a SOURCE.md with attribution to the original author.
Layout
skills/
├── <skill-name>/
│ ├── SKILL.md # the skill itself
│ ├── SETUP.md # (only if the skill needs machine side-effects) opt-in hook/cron install + reversal
│ └── SOURCE.md # (lifted skills only) attribution + license
Single-deep. No nesting, no plugins, no bundled binaries.
Most skills are pure markdown — making them visible is the install. A few need a hook or cron entry to work by default; those carry a SETUP.md next to their SKILL.md. The skill stays fully usable without it — SETUP.md only makes its behavior apply automatically.
Categories
Behavioral guardrails — keep model output honest and on-task
karpathy-guidelines— surgical changes, no over-engineeringcaveman— ultra-terse output modeanti-sycophancy— strip validating/hedging languageberzerk— relentless autonomous goal-pursuit mode (decide via /counsel, TDD, one thin slice at a time)
Hygiene — keep the codebase lean
trash-retired-files— reason-coded reversible GCdecompose-skill— split SKILL.md files > 100 linesprofiling-ladder— pick the right optimization rung (session > memory > skill > pipeline | TS > C > ASM)task-priority— UX > quality > security > scale > efficiencycodemap— deterministic (no-LLM) static-analysis snapshot → PlantUML + report: module shapes, seams, dead/untested/redundant code; diff before vs after a change
Engineering workflow (matt-pocock)
tdd,grill-with-docs,grill-me,zoom-outto-prd,to-issues,triage,prototype,diagnosehandoff,write-a-skillimprove-codebase-architecture,counselgit-guardrails-claude-code,setup-pre-commit,scaffold-exercises,migrate-to-shoehorn
Env-setup installers — write/install hooks instead of bundling runtime behavior
install-anti-sycophancy— writes Stop/UserPromptSubmit hookinstall-to-trash— writes PreToolUse hook replacingrmwith reversible trash-moveschedule-hygiene— writes cron/systemd-timer entries that fire hygiene skills viaclaude -p
Pre-PR
blog— drafts an entry in the ledger blog table via arc-agents API from the staged diff so the post is reviewed alongside the code
Model discovery
select-models— discovers reachable model providers, validates choices, writes~/.config/arc-skills.json
System design — reusable architecture frameworks
cam— Collector/Adaptor/Monitor judgment-gate framework: read wide, write narrow, measure over time, one append-only ledgerchampion-challenger— promote a challenger over the champion only past a pre-registered AND-gate
Session reflection (supply-chain / self-improvement)
dream— mine conversation history for failure modes, make one system improvementdream-insights— show latest dream journal without re-runningdream-status— show dream processing state and today's journal entry counts
Domain workflows — narrow vertical skills for specific external systems
vast-cli— drive thevastaiCLI correctly: search/create/poll/ssh/stop, with the proxy-unset + PATH-shim + ssh-rotation traps that silently no-op the CLIvast-instance— on-box best practices for vast.ai GPU jobs: bandwidth-gate,*-runtimeimages, the verified LFM2.5 pin stack, HF-token guard, stage-verify
Install
- Drop the
skills/directory (or any subset) into wherever your harness reads skills from. For Claude Code:~/.claude/skills/or a project's.claude/skills/. - Skills with a
SETUP.mdhave opt-in side-effects (hooks, cron). Run each skill interactively to trigger those side-effects — or skip the SETUP.md entirely and use the skill as pure markdown.
Skills without a SETUP.md need nothing beyond step 1.
Setting up on a fresh or non-Linux host (Windows/macOS)? Skills are pure markdown, so step 1 is the whole install — verify it worked by confirming your harness lists a skill from
skills/(e.g. it appears in Claude Code's/<skill-name>completion). If you run the doc-drift guard below, it's bash — use git-bash or WSL on Windows. Hit an install snag whose fix belongs in these steps? Open a PR so the next host doesn't hit it.
Curation principles
- Zero deps. A skill is a markdown file + optional small references. No installs.
- Distinct value. If two skills overlap, one of them goes.
- Genericized. No personal infra paths or private system refs in skill bodies.
- Attribution. Lifted skills carry
SOURCE.mdpointing to the original. - Light download, flexible install. Behavior that must live in the harness ships as an installer skill, not bundled code.
Doc-drift guard
Skill markdown can drift from the code it documents (wiring claims naming a
model that the agent file disagrees with, install commands that no longer
match the live install, private infra refs violating curation principle #3).
The self-heal loops (/dream, /token-waste, recency-gate) scope to the file
they identified and have not caught this drift shape in practice.
bin/arc-skills-doc-drift.sh is a grep test that fails on three known
falsified-claim patterns in skills/*/SKILL.md and skills/*/SETUP.md:
model: minimaxin skill markdown where the matchingagents/*.mdsaysmodel: haiku(the collector.md / SKILL.md wiring-claim drift).claude --bgin any SKILL.md / SETUP.md (the live install isclaude -p).home-lab-1in any SKILL.md / SETUP.md (private infra; curation principle #3).
Run it locally before opening a PR:
bin/arc-skills-doc-drift.sh
Add a new rule when a new drift shape appears — with a one-line comment naming the commit that introduced the pattern.