| name | hooks-creator |
| description | Use when the user asks to automate something on agent events (e.g. "run pdflatex after every edit", "open preview on save", "lint python after writes", "notify me when a tool fails"). Teaches you to propose a shell hook for user review via the propose_hook tool rather than editing .docwriter/hooks.json directly. |
hooks-creator
The workspace has a file at .docwriter/hooks.json that binds shell commands to Claude Agent SDK events. The user wants to manage these by asking you in natural language, then accepting or rejecting your proposal in the sidebar.
When to use this skill
The user describes an automation they want triggered by agent activity:
- "Run pdflatex after every Edit"
- "Lint the file I just saved"
- "Open Preview.app when I accept an edit"
- "Echo done after the agent finishes responding"
- "Notify me if a tool fails"
- "Snapshot a git commit after each response"
- "Format with Prettier after every Write"
Also use it when the user references build steps they want tied to edits (e.g. "recompile the manuscript", "rebuild the typst PDF").
How to propose a hook
Call the propose_hook tool exactly once per user request. Arguments:
- event (required): see the full list below.
- matcher (optional): regex over the tool name. Only meaningful for
PreToolUse, PostToolUse, and PostToolUseFailure. Common values:
"Edit|Write" — file mutations only"Edit" — edits only"Bash" — shell commands only- omit / empty — match every tool
- command (required): the shell command. Placeholders:
{{file}} — the edited file path (for tool-scoped events){{tool}} — the tool name
- reason (required): one sentence explaining what the hook does. Shown to the user on the proposal card.
Available events
Pick the event that matches when the user wants the command to run:
| Event | Fires when | Typical use |
|---|
PreToolUse | A tool is about to run | Validate input, short-circuit |
PostToolUse | A tool finished successfully (most common) | Build, lint, format, log |
PostToolUseFailure | A tool errored | Notifications, retries |
UserPromptSubmit | The user sends a new prompt | Start a timer, log the prompt |
Stop | The agent finishes its response | Compile, open preview, commit |
SubagentStop | A subagent (spawned via Task) finishes | Aggregate subagent results |
SessionStart | A new agent session begins | Log a banner, set up state |
SessionEnd | A session ends | Tear down state, flush logs |
Notification | SDK emits a notification (permission, idle, etc.) | Forward to Slack/PagerDuty |
Default to PostToolUse if the user's intent is unclear — it's the most common.
Examples
See examples/ next to this skill for reference shapes:
pdflatex-on-edit.json — PostToolUse + matcher, rebuild a PDF after editsruff-lint-edited-file.json — PostToolUse + {{file}} placeholdertypst-watch-stop.json — Stop event (no matcher)
Mirror the shape of the closest example when you propose. In particular, match the reason field style: short, single-sentence, describes what the user will observe.
DO NOT
- DO NOT write directly to
.docwriter/hooks.json via Edit/Write. The user wants to review hooks before they're saved.
- DO NOT propose multiple hooks per request unless the user explicitly asked for several. Ask clarifying questions first if ambiguous.
- DO NOT propose a hook whose command you're unsure about — ask the user to confirm the exact command instead of guessing.
- DO NOT set
matcher on non-tool events (Stop, UserPromptSubmit, Session*, Notification) — it's ignored.
After proposing
Keep your text response brief: "Proposed a hook that runs pdflatex after edits — accept in the sidebar to save it." The user reviews and accepts it from the OutlinePane.
