shared-skills

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #llm #android #deep-learning #linux #java #windows #vue #pytorch #openai #mcp

(★ 9)

No description available.

shared-skills Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# shared-skills

All Claude Code / OpenClaw skills in one place. Symlinked into projects via `setup.sh`.

## Structure

```
skills/           # Skill definitions (SKILL.md per skill)
hooks/user-level/ # Claude Code hooks (symlinked to ~/.claude/hooks/)
hooks/copilot/    # Copilot CLI hook configs (symlinked to ~/.copilot/hooks/)
project-hooks/    # Project-level hook systems (installed per-repo, not symlinked)
scripts/          # Helper scripts used by setup.sh
setup.sh          # Symlinks skills + hooks, registers hooks in settings.json
```

## Setup

```bash
# Local (Claude Code + Copilot CLI) — symlinks skills + hooks, registers hooks in settings.json
./setup.sh local

# Preview what would change
./setup.sh local --dry-run

# VPS (OpenClaw) — pulls repo + creates symlinks
./setup.sh openclaw
```

## Skill inventory

### User-level (all projects)

Symlinked into **both** `~/.claude/skills/` (Claude Code) and `~/.copilot/skills/`
(Copilot CLI — it does not read `~/.claude/skills/`). Project-level skills need no
mirror: Copilot CLI reads `.claude/skills/` directly.

| Skill | Category |
|-------|----------|
| ai-agent-debugging-guide | Engineering patterns |
| async-dedup-race-in-streaming-pipelines | Engineering patterns |
| fix-at-appropriate-layer | Engineering patterns |
| git-commit-practices | Engineering patterns |
| test-debugging-without-hacking | Engineering patterns |
| eli5 | Writing |
| agent-team-orchestration-patterns | Agent/orchestration |
| bug-basher-5000 | Agent/orchestration |
| create-cli-skill | Agent/orchestration |
| claude-code-session-transcript-analysis | Agent/orchestration |
| claude-code-token-usage | Agent/orchestration |
| claudeception | Agent/orchestration |
| code-review | Agent/orchestration |
| ralph-wiggum-loop-setup | Agent/orchestration |
| simplify | Agent/orchestration |
| browser-tab-resource-investigation | macOS/system |
| playwright-cli | macOS/system |
| swift-macos-native-apis | macOS/system |
| speech-to-text | macOS/system |
| gog | Google tools |
| google-calendar-ics-timezone-handling | Google tools |
| notion-child-page-preservation | Notion |
| notion-page-link-syntax | Notion |
| git-worktree-nextjs-prisma | Next.js/Prisma |
| nextjs-client-server-boundary-dns-error | Next.js/Prisma |
| nextjs-prisma-client-component-import | Next.js/Prisma |
| prisma-7-driver-adapter | Next.js/Prisma |
| prisma-migration-drift-recovery | Next.js/Prisma |
| vitest-class-mocking | Vitest |
| vitest-mock-implementation-persistence | Vitest |
| vitest-recharts-mocking | Vitest |

### Project: personal_assistant_claude

| Skill | Purpose |
|-------|---------|
| beeper | Cross-platform messaging CLI |
| femi-calendar | Google Calendar management |
| macos-messages-contacts-access | macOS Messages/Contacts access |
| message-review | Scan messages for action items |
| notion-style-matching | Match Notion formatting preferences |
| notion-todo-organization | Organize Notion todo database |
| querying-and-updating-notion | Notion MCP vs API decision routing |
| swarm-research | Multi-agent research with varied context |

### Project: peronal_budget_tracking

| Skill | Purpose |
|-------|---------|
| architecture-parity-review | Architecture consistency checks |
| multi-agent-orchestration | Multi-agent coordination patterns |

## Readonly Gate Hook

The **safely-skip-permissions** hook (`readonly-gate.sh`) auto-approves read-only Claude Code tool calls (file reads, searches, git status, etc.) so you aren't prompted for every safe operation. Writes still require confirmation.

**Install (standalone — no need to clone the full repo):**

```bash
# macOS / Linux
git clone https://github.com/DavidTeju/shared-skills.git /tmp/shared-skills
bash /tmp/shared-skills/hooks/user-level/install-readonly-gate.sh
rm -rf /tmp/shared-skills
```

```powershell
# Windows (PowerShell) — requires Perl (e.g. Strawberry Perl) on PATH
git clone https://github.com/DavidTeju/shared-skills.git $env:TEMP\shared-skills
& $env:TEMP\shared-skills\hooks\user-level\install-readonly-gate.ps1
Remove-Item -Recurse -Force $env:TEMP\shared-skills
```

This copies the hook to `~/.claude/hooks/` and registers it in `~/.claude/settings.json`. Run the install script again anytime to update to the latest version.

To preview what it would do: `bash install-readonly-gate.sh --dry-run` (or `.\install-readonly-gate.ps1 -DryRun` on Windows)

Unlike `settings.json`'s config which uses a simple allowlist, this is more intelligent. It's a two-phase bash classifier. Named tools (Read, Grep, Glob, etc.) are checked against an allowlist, but Bash commands go through a pipeline-aware parser that splits on `|`, `&&`, `;`, and `||`, then classifies each segment individually. It handles quoted strings, `$()` command substitutions, environment variable prefixes, dangerous flags on otherwise-safe commands (e.g. `find -exec`, `sed -i`, `awk system()`), output redirects vs. safe `/dev/null` redirects, and multi-word read patterns like `git log`, `gh pr list`, and `gog calendar events`. When in doubt, it defers to the normal permission prompt. The principle is: false negatives are safe, false positives are not.

## Command Blocker Hook

`hooks/user-level/command-blocker.sh` is a `preToolUse` hook that denies banned shell commands and
tells the agent what to do instead. It is a **rule registry**, not a single-purpose blocker — shared
plumbing parses the agent's command into individual invocations, and each rule inspects one.

| Rule | Blocks | Tells the agent to |
|------|--------|--------------------|
| `rule_rm_rf` | recursive **and** force `rm` | use `trash` (macOS Trash is recoverable; `rm -rf` is not) |
| `rule_git_blind_staging` | `git add -A/--all/./-u/*/:/`, `git stage` sweeps, `git commit -a/--all` | run `git status --short`, stage explicit paths, then **read** `git diff --staged` before committing |

Blind staging is banned because the working tree is shared with the human and other agents: a sweep
silently commits their uncommitted work, secrets, and scratch files, and no accurate commit message
can be written for a diff nobody read.

It runs on **both agents** from one script:

| Agent | Config | Payload dialect |
|-------|--------|-----------------|
| Claude Code | `~/.claude/settings.json` → `PreToolUse`, matcher `Bash` | PascalCase (`tool_name` / `tool_input`) |
| Copilot CLI | `~/.copilot/hooks/command-blocker.json` → `preToolUse`, matcher `bash\|shell` | camelCase (`toolName` / `toolArgs`) |

The script reads either payload shape and emits a deny in both dialects at once — top-level
`permissionDecision` for Copilot, nested `hookSpecificOutput` for Claude Code.

**Detection** tokenizes each line with `shlex` and only flags commands in *command position*, so
quoted mentions are left alone. Heredoc bodies are stripped before parsing — they are data, so a
commit message or doc that quotes a banned command does not trip a rule (`<<<` here-strings are not
heredocs and are left intact). It sees through chaining (`&&`, `;`, `|`, newlines), absolute paths
(`/bin/rm`, `/usr/bin/git`), env prefixes (`FOO=1 …`), wrappers (`sudo`, `xargs`, `env`, `nohup`, …),
and git's global options (`git -C /repo add -A`). It does **not** flag `rm -f`, `rm -r`, `npx rimraf`,
`git add <path>`, `git add -p`, `git commit -m`, `git submodule add`, or `grep -r 'rm -rf' .`.

It always exits `0`. Copilot treats a non-zero exit from a `preToolUse` command hook as a deny, so a
crash in this script must not block unrelated tool calls. **That guarantee covers runtime errors
only** — a bash *syntax* error stops the script before `exit 0` and denies every shell command on the
machine. The Python payload is assigned inside single quotes, so it must never contain a literal
apostrophe (use `\x27`); the test suite runs `bash -n` first so this fails in tests instead of in
production.

**Adding a rule:** add cases to `command-blocker.test.sh` first and watch them fail, then write a
`rule_*(base, args) -> str | None` function and append it to `RULES`.

**Renaming a hook:** add the old filename to `RETIRED_HOOKS` in `setup.sh`. A dangling symlink from a
rename makes the hook error, and an erroring `preToolUse` hook denies *every* shell command. Agent
sessions cache hook config at startup, so restart running sessions after a rename.

Install: `./setup.sh local`. Tests: `bash hooks/user-level/command-blocker.test.sh -v` (202 cases).

## Project-level hooks

Unlike the user-level hooks above (symlinked into `~/.claude/hooks/` for every project),
`project-hooks/` holds hook systems that install **into a specific repo's** `.claude/` and
`.github/`. They are not touched by `setup.sh` — each ships its own installer.

### skill-gate

[`project-hooks/skill-gate/`](project-hooks/skill-gate/) makes sure the **right skill is loaded
before a gated workflow runs** — reminding on matching prompts and blocking gated commands/MCP calls
until the skill is loaded. One shared engine drives both **Claude Code** and the **GitHub Copilot
CLI**. The shipped gates cover skills in this repo (`git-commit-practices`, `code-review`,
`playwright-cli`, `update-readme`, `create-cli-skill`); edit `skill-gates.json` to fit any repo.

**Install into a repo (standalone):**

```bash
git clone https://github.com/DavidTeju/shared-skills.git /tmp/shared-skills
node /tmp/shared-skills/project-hooks/skill-gate/install.mjs /path/to/repo   # add --dry-run to preview
rm -rf /tmp/shared-skills
```

The installer copies the runtime files and registers the hooks in the target repo's
`.claude/settings.json` and `.github/hooks/hooks.json` (idempotent). See the
[skill-gate README](project-hooks/skill-gate/README.md) for gates, schema, and tests.

## Adding a new skill

1. Create `skills/<name>/SKILL.md`
2. Add it to the appropriate list in `setup.sh`
3. Run `./setup.sh local` to symlink
4. `git push` then `./setup.sh openclaw` to deploy to VPS

shared-skills

shared-skills Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

shared-skills

shared-skills Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README