convez.ai
SkillAutomaçãoIntermediário
dr-snob

BBR Claude Skills

dr-snob/bbr-claude-skills

Skills reutilizáveis para Claude Code — orquestração, contexto e padrões de delegação

Ver no GitHub
InicialPluginScore · de 0 a 100

Coleção comunitária de skills para Claude Code que ensinam o modelo a lidar melhor com tarefas específicas. Inclui orquestração de subagents, preservação de contexto e protocolos de time multi-agente, com instalação global ou por projeto.

Instalação

$npx convez-ai dr-snob/bbr-claude-skills

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
57%
Último commit
há 3 semanas
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Orquestração paralela de subagentes com verificação de build
  • Preservação automática de contexto e memória em sessões longas
  • Coordenação de times multi-agente com votação por consenso
  • Otimização de fluxo de trabalho em projetos complexos

Veredito

Conjunto prático de skills que reduz overhead de coordenação em workflows com Claude Code. Bem estruturado, segue padrão aberto e permite reutilização em equipes via git — ideal para quem trabalha com múltiplos agentes ou contextos longos.

README

A tradução para o português deste README ainda está na fila.

BBR Claude Skills

Community-driven skill collection for Claude Code — Anthropic's CLI for Claude.

Skills teach Claude Code how to handle specific tasks better. They load on-demand, preserve context, and work across all your projects.

What's Inside

SkillDescription
subagent-orchestrationParallel subagent deployment with build verification, context priming, memory extraction, and 6 core patterns for efficient delegation.
context-sentinelAutomatic session context preservation — persists plans, progress, and decisions to memory files. Includes a pressure detection hook that monitors context window usage and triggers checkpoints before compaction.
house-party-protocolMulti-agent team orchestration with study group swarming, model-per-role strategy, consensus voting, and hook-based quality gates.

Contribute yours!

Install

npx bbr-claude-skills

That's it. You'll be asked to choose a scope:

  Where should skills be installed?

  1  Global (~/.claude/skills/)
     Available in all your projects

  2  Project (.claude/skills/)
     Only this project, can commit to git for team sharing

  Choose [1/2]:

Update

npx bbr-claude-skills update

Other Commands

CommandWhat it does
npx bbr-claude-skillsInteractive install with scope selection
npx bbr-claude-skills updatePull latest from GitHub
npx bbr-claude-skills listShow installed skills
npx bbr-claude-skills versionShow version info
npx bbr-claude-skills uninstallRemove all managed skills

Usage in Claude Code

Automatic

Claude Code automatically discovers installed skills and uses them when relevant. Just work normally — Claude will apply the right skill when a task matches.

Manual Invocation

Invoke any skill directly with its slash command:

/subagent-orchestration

How It Works

Skills follow the Agent Skills open standard adopted by Anthropic.

GitHub Repo                        Your Machine
────────────                       ────────────
skills/                            ~/.bbr-claude-skills/repo/  (cloned)
  subagent-orchestration/              ↓ symlinked
    SKILL.md              ───→     ~/.claude/skills/subagent-orchestration/
                                   (or .claude/skills/ for project scope)

The installer clones the repo once, then creates symlinks. Updates pull latest and re-link — no files copied, always up to date.

Each skill has a SKILL.md with:

  • YAML frontmatter — name and description for discovery
  • Markdown body — Instructions Claude follows when the skill activates

Included Skills

Subagent Orchestration

Efficient patterns for deploying subagents to maximize throughput and preserve context.

6 Core Patterns:

PatternWhen to UseAgent Type
Context PrimingBefore implementing a featureExplore (background)
Parallel ResearchMultiple independent lookupsExplore (background)
Memory ExtractionContext window filling upgeneral-purpose (background)
Server Log MonitoringTesting with local serverBash background task
Background Test WritingFeature + tests simultaneouslygeneral-purpose (background)
Isolating Verbose OpsRunning test suites/buildsgeneral-purpose (foreground)

Build Verification Rule (added Feb 2026): Any subagent that modifies code MUST verify the build passes before reporting done. This prevents cascading type errors from unverified changes — learned from a real session where a type-fixing agent introduced 8 build errors.

MANDATORY: After all edits, run the project build command and fix any errors before reporting done.
Build command: cd /path/to/project && rm -rf .next && npm run build
If the build fails, fix the errors and rebuild until green. Do NOT report completion with a failing build.

Core principle: Your main conversation is expensive real estate. Delegate verbose, exploratory, or independent work to subagents.


Context Sentinel

Automatic context preservation that ensures no work is lost when context compacts or sessions end.

The Problem: Claude Code auto-compacts context at arbitrary points, often mid-task. Plans, decisions, and progress are lost — forcing expensive reconstruction on resumption.

The Solution: Two-role architecture (Scribe + Curator) that persists session state to files:

RoleModelWhenWhat It Does
ScribeHaikuAfter each taskUpdates session-live.md with plan progress, decisions, errors
CuratorSonnetEvery 3-4 checkpointsExtracts durable learnings to MEMORY.md, trims session file

Context Pressure Detection via pressure-check.sh hook:

LevelSignalAction
GREEN< 500KB transcript, < 25 tool callsNormal checkpointing
YELLOW500KB-1MB, 25-50 toolsIncrease checkpoint frequency
RED> 1MB, 50+ toolsImmediate checkpoint, warn user
EMERGENCY> 1.5MB, 70+ toolsFull dump, recommend session restart

Hook Setup (add to ~/.claude/settings.json):

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Edit|Write|Bash",
        "hooks": [
          {
            "type": "command",
            "command": "~/.claude/skills/context-sentinel/pressure-check.sh"
          }
        ]
      }
    ]
  }
}

Planning Mode Handling:

ScenarioWhat Happens
Plan created successfullyScribe records full plan to session-live.md
Context compacts mid-planPartial plan saved with exploration progress
Session ends mid-executionPlan + progress + current task preserved
Pressure spikes during planningStop exploring, synthesize from what's known

Resumption: Next session reads session-live.md first. Instantly knows what was planned, what's done, and what's next — no re-derivation needed.

File Structure:

.claude/projects/{project}/memory/
  MEMORY.md          # Durable learnings (survives across sessions)
  session-live.md    # Active session state (ephemeral, structured)

Core principle: The cheapest token is one you never have to spend twice.


House Party Protocol

Multi-agent team orchestration that goes beyond vanilla agent teams:

FeatureWhat It Adds
Study Group SwarmingFree agents join busy ones — guided subtask splitting, not random task claiming
Model-per-RoleOpus for specialists, Sonnet for workers, Haiku for scouts
Consensus VotingAdversarial debate, parallel+judge, majority agreement
Quality GatesTeammateIdle and TaskCompleted hook enforcement
Party CompositionsPre-built team templates (Research, Build, Debug, Review) with model assignments

Core principle: Better orchestration beats smarter models. Break work into focused subtasks, let agents solve them in parallel, and when someone finishes — they join whoever needs help most.

Project Structure

bbr-claude-skills/
├── README.md
├── LICENSE                 # MIT
├── CONTRIBUTING.md         # How to add skills
├── package.json            # npm package (enables npx)
├── bin/
│   └── cli.js              # CLI entrypoint
├── skills/
│   ├── subagent-orchestration/
│   │   └── SKILL.md        # 6 patterns + build verification rule
│   ├── context-sentinel/
│   │   ├── SKILL.md         # Two-role architecture + pressure detection
│   │   └── pressure-check.sh  # PreToolUse hook script
│   └── house-party-protocol/
│       └── SKILL.md
└── docs/
    └── skill-authoring.md  # Guide for creating new skills

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

To add a skill:

  1. Fork this repo
  2. Create skills/your-skill-name/SKILL.md
  3. Test locally with ln -s $(pwd)/skills/your-skill-name ~/.claude/skills/your-skill-name
  4. Open a Pull Request

Resources

License

MIT — see LICENSE


Built by BlackBodyRad with Claude Code.

Ver o original no GitHub

Bom para

Relacionados