e8e572e0af
New skills for container agent task orchestration: - /decompose: break tasks into subtasks with dependency graph, write .agent-tasks.json - /orchestrate: check task state, launch container agents in git worktrees, poll for completion Updated with learnings from first real run (agent-runtimes M3, 9 tasks): - Prompts must end with "run pytest and fix failures" - State import conventions explicitly in prompts - Warn about uncommitted files before decomposing (worktrees need committed content) - Copy dependency outputs into downstream worktrees before launching Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
53 lines
3.8 KiB
Markdown
53 lines
3.8 KiB
Markdown
# Custom Claude Skills
|
|
|
|
Reusable Claude Code skills shared across all projects via symlinks into `~/.claude/skills/`.
|
|
|
|
## Repository Structure
|
|
|
|
```
|
|
custom-claude-skills/
|
|
├── CLAUDE.md # This file
|
|
├── MEMORY.md # Learnings about skill development
|
|
├── FUTURE.md # Future skill ideas
|
|
├── README.md # Setup instructions and skill catalogue
|
|
├── scripts/
|
|
│ └── install.sh # Symlinks all skills into ~/.claude/skills/
|
|
└── skills/
|
|
├── context-load/SKILL.md # Reload project context
|
|
├── decompose/SKILL.md # Task decomposition for container agents
|
|
├── distill-best-practices/SKILL.md # Cross-project best practices
|
|
├── housekeeping/SKILL.md # Cross-project health check
|
|
├── linter/SKILL.md # Format/lint audit and execution
|
|
├── log/SKILL.md # End-of-session logging
|
|
├── orchestrate/SKILL.md # Container agent dispatch with git worktrees
|
|
├── reflect/SKILL.md # Milestone reflection
|
|
└── reflect-logs/SKILL.md # Process logs into memory
|
|
```
|
|
|
|
## Conventions
|
|
|
|
- Each skill lives in `skills/<skill-name>/SKILL.md`
|
|
- Skills should be project-agnostic where possible — use dynamic context injection (`!`command``) to adapt to the current project
|
|
- The `scripts/install.sh` script creates symlinks from `~/.claude/skills/<name>` → this repo's `skills/<name>/`
|
|
- After adding a new skill, run `./scripts/install.sh` to register it
|
|
- Skills that are only useful for one project should live in that project's `.claude/skills/` instead
|
|
- Every skill must have a corresponding `memory/skill-<name>.md` in `claude-foundations` with purpose, usage, how it works, and gotchas — update the `claude-foundations/MEMORY.md` index as part of the same commit
|
|
|
|
## Skill Development Guidelines
|
|
|
|
- **Inline by default** — only use `context: fork` if the skill genuinely doesn't need conversation history
|
|
- **Pre-fetch context** with `!`command`` injection to reduce tool calls during execution
|
|
- **Restrict tools** with `allowed-tools` to the minimum needed — this reduces permission prompts
|
|
- **Use $ARGUMENTS** for user input, `$0`, `$1` etc. for positional args
|
|
- Dynamic commands in `!`command`` run at skill load time, not during Claude's execution
|
|
|
|
## `!`command`` Gotchas
|
|
|
|
The `!`command`` syntax in SKILL.md runs shell commands at skill load time to pre-gather context. These commands go through Claude Code's Bash permission checker against the `allowed-tools` patterns. Key constraints:
|
|
|
|
- **No `$()` command substitution** — the permission checker rejects commands containing `$()` even if the outer command matches an allowed pattern. This means you cannot nest subshells (e.g., `git log --since="$(git log ...)"` will fail).
|
|
- **No complex shell pipelines relying on subshells** — keep `!`command`` commands simple and self-contained. If you need complex logic, have the skill instructions tell Claude to run it via tool calls instead.
|
|
- **`allowed-tools` patterns must match the command binary** — `Bash(git *)` allows any `git` subcommand, but `Bash(cat *)`, `Bash(sed *)`, `Bash(ls *)` must be listed separately if those commands appear in `!`command`` blocks.
|
|
- **Prefer specific tool patterns over broad ones** — `Bash(git log *)`, `Bash(git diff *)` is safer than `Bash(git *)` to avoid accidentally allowing destructive git commands (e.g., `git reset`, `git push --force`).
|
|
- **Fallback to tool instructions for dynamic paths** — if a command needs `$ARGUMENTS` to compute a file path (e.g., `cat scripts/verify-m${N}.sh`), replace the `!`command`` with a plain-text instruction telling Claude to use the Read tool. This avoids `$()` substitution and is more reliable.
|