claude-prompt
Hbrinj/claude-prompt
Centralised collection of reusable Claude Code agents, shared across multiple repositories
Instalação
npx convez-ai Hbrinj/claude-prompt1 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
README
A tradução para o português deste README ainda está na fila.
Claude Agents & Skills
A centralised collection of reusable Claude Code agents and skills, designed to be shared across multiple repositories.
Overview
This repository contains custom Claude Code subagents and slash-command skills that can be installed globally or per-project. Each agent and skill is a self-contained Markdown file.
Structure
agents/
├── README.md # Agent catalogue and usage notes
└── *.md # Individual agent definitions
skills/
├── README.md # Skills catalogue and usage notes
└── *.md # Individual skill definitions
SYSTEM_PROMPT.md # Coordinator workflow skeleton — source of truth with skills/implement-feature.md (committed)
CLAUDE.md # Repo conventions for Claude when working in this repo
Workflow source of truth
The canonical, version-controlled coordinator workflow is split across two files: SYSTEM_PROMPT.md (always-loaded skeleton — role, non-negotiable rules, Step 1 → Understand/Specify/Slice and Step 2 → Implement outline) and skills/implement-feature.md (Step 2 procedure, agent/skill routing index, checkpoint format, feature-log schema — loaded on demand). Two derived copies of SYSTEM_PROMPT.md exist downstream:
- Consumer repos'
CLAUDE.md— kept in sync automatically by thesync-upstreamskill (skills/sync-upstream.md), which writes the latestSYSTEM_PROMPT.mdcontent into a guarded block bounded by<!-- SYSTEM_PROMPT:START -->/<!-- SYSTEM_PROMPT:END -->markers. Anything in the consumer repo'sCLAUDE.mdoutside those markers is preserved. - Your
~/.claude/CLAUDE.md— Claude Code's global instructions file. Kept in sync manually —sync-upstreamdoes NOT touch it. After editingSYSTEM_PROMPT.md, copy the new content into~/.claude/CLAUDE.mdyourself if you want the active session and other "raw" sessions (no sync-upstream installed) to pick up the change immediately.
When changing the coordinator workflow, ALWAYS edit
SYSTEM_PROMPT.mdand/orskills/implement-feature.mdin this repo — NEVER edit a derived copy (~/.claude/CLAUDE.mdor any consumer repo'sCLAUDE.md) directly.Editing a derived copy alone is upside-down: the change is not version-controlled, is not shared with other machines or users, and is silently overwritten the next time
sync-upstreamruns in that consumer repo. The fix is one-way: editSYSTEM_PROMPT.mdhere, commit, then propagate (sync-upstream for consumer repos, manual copy for~/.claude/CLAUDE.md).If you're unsure where a coordinator-shaped change belongs, default to the source-of-truth pair:
SYSTEM_PROMPT.mdfor skeleton/rules changes,skills/implement-feature.mdfor Step 2 procedure changes.
The same source-of-truth rule applies to agents/*.md and skills/*.md — the repo is the source; ~/.claude/agents and ~/.claude/commands are derived (typically symlinked by sync-upstream, see Usage).
Usage
Global install (available in every repo)
# Symlink or copy individual agents
cp agents/<agent-name>.md ~/.claude/agents/
# Or symlink the whole directory (changes here apply everywhere)
ln -s "$(pwd)/agents" ~/.claude/agents
Per-project install
mkdir -p .claude/agents
cp path/to/claude-prompt/agents/<agent-name>.md .claude/agents/
Invoke an agent
Once installed, reference any agent in Claude Code:
Use the <agent-name> agent to …
Or from the CLI:
claude --agent <agent-name> "your task here"
Skills
Skills are reusable slash commands (/skill-name) that appear in Claude Code's command palette.
Global install
cp skills/<skill-name>.md ~/.claude/commands/
# Or symlink the whole directory
ln -s "$(pwd)/skills" ~/.claude/commands
Per-project install
mkdir -p .claude/commands
cp path/to/claude-prompt/skills/<skill-name>.md .claude/commands/
Invoke a skill
/skill-name
Adding a new skill
- Create a new
.mdfile inskills/. - Write the prompt body (no frontmatter required, but
description:is recommended). - Open a PR — the skill becomes available to all repos on merge.
- Update
skills/README.mdwith a one-line entry.
Adding a new agent
- Create a new
.mdfile inagents/. - Add the required frontmatter (
name,description,tools). - Write the system prompt body.
- Open a PR — the agent becomes available to all repos on merge.
Agent template
---
name: agent-name
description: One-line description of when Claude should use this agent.
tools: Read, Edit, Bash # comma-separated list of tools this agent may use
---
You are a specialist in …
## Responsibilities
- …
## Constraints
- …
Guard hooks
scripts/hooks/ contains PreToolUse hooks that mechanically enforce two workflow non-negotiables: guard-main-edit.sh denies Edit/Write/NotebookEdit while the target's repo is on main/master, and guard-push-main.sh denies git push to main/master. install.sh --global symlinks ~/.claude/hooks → <clone>/scripts/hooks and prints the settings.json block to add manually (it never edits settings.json). Both hooks require jq (preinstalled on recent macOS; otherwise brew install jq) and fail open when jq is missing or stdin is malformed. Per-path opt-outs for the edit guard live in ~/.claude/hooks-exceptions. Full contract: scripts/hooks/README.md.
Contributing
- Keep each agent and skill focused on a single responsibility.
- For agents: prefer narrow tool lists — only grant the tools an agent genuinely needs.
- Update
agents/README.mdorskills/README.mdwith a one-line entry for every new addition.