Why Use This This skill provides specialized capabilities for aiskillstore's codebase.
Use Cases Developing new features in the aiskillstore repository Refactoring existing code to follow aiskillstore standards Understanding and working with aiskillstore's codebase structure
Install Guide 2 steps 1 2 Install inside Ananke
Click Install Skill, paste the link below, then press Install.
https://github.com/aiskillstore/marketplace/tree/main/skills/brendanshields/creating-hooks Skill Snapshot Auto scan of skill assets. Informational only.
Valid SKILL.md Checks against SKILL.md specification
Source & Community
Updated At Jan 19, 2026, 04:39 AM
Skill Stats
SKILL.md 187 Lines
Total Files 1
Total Size 0 B
License NOASSERTION
---
name: creating-hooks
description: |
Creates Claude Code hooks for automation and workflow customization.
Guides through hook events, configuration, and script creation.
Use when user wants to create a hook, automate Claude Code, or asks about hook events.
---
# Creating Hooks
Guides creation of Claude Code hooks for automation and workflow customization.
## Quick Start
1. Choose hook event (when should it trigger?)
2. Configure in settings.json
3. Create hook script
4. Test the hook
## Workflow: Create New Hook
```
Progress:
- [ ] Select hook event
- [ ] Add to settings.json
- [ ] Create hook script
- [ ] Test and validate
```
### Step 1: Select Hook Event
| Event | When It Triggers | Common Use |
|-------|------------------|------------|
| `PreToolUse` | Before tool runs | Block/modify tools |
| `PostToolUse` | After tool succeeds | Validate, log, feedback |
| `UserPromptSubmit` | User sends message | Inject context, validate |
| `SessionStart` | Session begins | Load context, init state |
| `SessionEnd` | Session ends | Cleanup, save state |
| `Stop` | Agent finishes | Decide if should continue |
Full event reference: [reference.md](reference.md)
### Step 2: Configure settings.json
Location priority (highest wins):
1. `.claude/settings.local.json` (local, not committed)
2. `.claude/settings.json` (project)
3. `~/.claude/settings.json` (user)
Basic structure:
```json
{
"hooks": {
"EventName": [
{
"matcher": "ToolPattern",
"hooks": [
{
"type": "command",
"command": "bash \"$CLAUDE_PROJECT_DIR/.claude/hooks/my-hook.sh\""
}
]
}
]
}
}
```
### Step 3: Create Hook Script
Use templates from [templates/](templates/) directory.
Key requirements:
- Read JSON from stdin
- Use exit codes for control (0=success, 2=block)
- Output JSON for decisions
### Step 4: Test
Run hook manually with test input:
```bash
echo '{"tool_name":"Write"}' | bash .claude/hooks/my-hook.sh
```
## Hook Configuration
### Matcher Patterns
```json
"matcher": "Write" // Exact match
"matcher": "Edit|Write" // Multiple tools
"matcher": "mcp__.*" // MCP tools (regex)
"matcher": "*" // All tools
```
Matchers apply to: `PreToolUse`, `PostToolUse`, `PermissionRequest`
### Timeout
```json
{
"type": "command",
"command": "...",
"timeout": 120
}
```
Default: 60 seconds. Max recommended: 300 seconds.
## Exit Codes
| Code | Meaning | Behavior |
|------|---------|----------|
| 0 | Success | Continue normally |
| 2 | Block | Stop action, show error |
| Other | Non-blocking error | Log only (verbose mode) |
## JSON Output
Return JSON to stdout for decisions:
```json
{
"decision": "block",
"reason": "Why blocked",
"additionalContext": "Info for Claude"
}
```
Decision values by event:
- `PreToolUse`: `allow`, `deny`, `ask`
- `PostToolUse`: `block` (with reason)
- `UserPromptSubmit`: `block` (with reason)
- `Stop`: `block` (requires reason)
## Security Best Practices
1. **Quote all variables**: `"$VAR"` not `$VAR`
2. **Use absolute paths**: `"$CLAUDE_PROJECT_DIR/..."`
3. **Validate inputs**: Check before processing
4. **Block path traversal**: Reject paths with `..`
5. **Set timeouts**: Prevent runaway scripts
## Environment Variables
Available in all hooks:
- `CLAUDE_PROJECT_DIR` - Project root path
- `CLAUDE_CODE_REMOTE` - "true" if web environment
SessionStart only:
- `CLAUDE_ENV_FILE` - Path to persist env vars
## Common Patterns
### Inject Context on Session Start
```bash
#!/bin/bash
# Output context for Claude
echo '{"additionalContext": "Project uses TypeScript"}'
exit 0
```
### Block Dangerous File Edits
```bash
#!/bin/bash
INPUT=$(cat)
FILE=$(echo "$INPUT" | jq -r '.tool_input.file_path // empty')
if [[ "$FILE" == *".env"* ]]; then
echo "Blocking edit to sensitive file" >&2
exit 2
fi
exit 0
```
### Log All Tool Usage
```bash
#!/bin/bash
INPUT=$(cat)
TOOL=$(echo "$INPUT" | jq -r '.tool_name')
echo "$(date -Iseconds) $TOOL" >> "$CLAUDE_PROJECT_DIR/.claude/tool.log"
exit 0
```
See [reference.md](reference.md) for complete event details and more examples.