Skills — Disciplinas Reutilizáveis para Agentes Claude
MrBinnacle/skills
Disciplinas de verificação e decisão reutilizáveis para agentes Claude confiáveis
Ver no GitHubColeção de skills compostos que implementam disciplinas de verificação, qualidade de decisão e ritmo de engenharia para agentes Claude. Cada skill é um módulo standalone com checklist de modos de falha documentados (gotchas.md) para aumentar confiabilidade em julgamentos críticos. Projetado para impedir que confiança do agente substitua verificações reais.
Instalação
npx convez-ai MrBinnacle/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
- Implementar checkpoints de verificação antes de assertions críticas em agentes
- Orquestração de workflows com transições closure→build validadas
- Separar lógica de decisão da ratificação com múltiplos SMEs
Veredito

Recurso técnico de alta qualidade para construir agentes mais confiáveis através de disciplinas verificáveis. Melhor para quem trabalha com sistemas críticos onde confiança não pode vir apenas de confiança do modelo.
README
A tradução para o português deste README ainda está na fila.
skills
Confidence is not evidence.
Most agent failures at the judgment layer are not reasoning failures or missing context. They are verification failures: the agent's confidence stood in for a check that never happened. These are composable disciplines that put the check back in the loop — verify before asserting, separate a decision from its ratification, distrust an instrument that disagrees with itself, log each failure mode so the next run does not relearn it.
Documentation and context frameworks keep an agent oriented to where the work goes. These skills govern whether its judgment can be trusted to get there: orchestration, decision quality, and engineering rhythm, each a standalone discipline you can drop into Claude Code or any compatible runtime.
Every skill carries its own append-only gotchas.md: the observed and
anticipated failure modes that produced the discipline.
Authored by Matthew Gruber. Layout inspired by github.com/mattpocock/skills.
Available skills
Engineering — workflow disciplines for shipping software
- closure-mode-at-boundaries — codifies the closure→build transition at sprint/phase boundaries. Dispatch a parallel SME swarm, then execute the swarm's action list before presenting a revised frame. Prevents the failure mode of forwarding multi-voice menus instead of executed verifications.
- azimuth — placeholder; added separately.
Install
Claude Code
Each skill is a directory containing SKILL.md + sibling files. Claude Code discovers skills by directory name; the SKILL.md frontmatter description: field drives auto-invocation.
Option 1 — clone the whole collection:
git clone https://github.com/MrBinnacle/skills.git ~/.claude/skills/mr-skills
Then symlink the skill directories you want into your skills root:
ln -s ~/.claude/skills/mr-skills/skills/engineering/closure-mode-at-boundaries \
~/.claude/skills/closure-mode-at-boundaries
Option 2 — copy individual skill directories:
git clone https://github.com/MrBinnacle/skills.git /tmp/mr-skills
cp -r /tmp/mr-skills/skills/engineering/closure-mode-at-boundaries \
~/.claude/skills/
Other Claude-compatible runtimes
Each skill's SKILL.md uses the Anthropic Agent Skills frontmatter convention (name: + description:). Runtimes that follow the same convention should discover the skills with their native mechanism. Per-skill prerequisites.md files list required runtime capabilities (parallel dispatch, ≥2 subagents, etc.).
Repository layout
skills/
engineering/ ← workflow disciplines for shipping software
README.md bucket index
closure-mode-at-boundaries/
SKILL.md entry point
swarm-composition.md roster + role-to-runtime mapping
transition.md 5-step closure→build discipline
case-study.md worked example
formalization.md wire into your lock-skill terminal step
prerequisites.md runtime + project surface requirements
prompt-templates.md copy-pasteable per-role prompts
gotchas.md append-only failure-mode log
azimuth/
README.md placeholder
Authoring conventions
If you want to contribute or adapt:
- Frontmatter is minimal —
name:+description:only. Description ≤ 200 chars. Triggers baked into the description sentence ("Use when X, Y, Z"). - Naming convention —
UPPERCASE-NAMED.mdfor documents / templates / formats.lowercase-named.mdfor concepts / aspects / principles.SKILL.mdis always uppercase. - Sizes —
SKILL.md400 B to ~7 KB. Aux files 400 B to ~3 KB each. IfSKILL.mdis over 5 KB, you are probably bundling too much — split into sibling aux files. - Cross-references — inline at moment-of-need. No trailing "Related" / "See also" section.
- Every skill ships
gotchas.md— append-only log of OBSERVED + ANTICIPATED failure modes. Never delete entries. - Discipline vs implementation — make explicit which parts of the skill are the stable contract (the discipline) vs. illustrative (specific subagent IDs, paths, project names). Adopters need to know what they can swap.
Contributing
Issues and PRs welcome. Each skill is independently versioned via the repo's git history. New skills:
- Pick the appropriate bucket directory (or propose a new one).
- Add the skill directory under the bucket.
- Update the bucket's
README.mdto list the skill with a one-line description linking to itsSKILL.md. - Update the top-level
README.mdto list the skill under its bucket.
License
MIT — see LICENSE.