Exploring automatic invocation after installation

[HeatCrab]

After installing and using this project for a while, I have a question. My understanding is that the project is heading in the direction of a linter, but as a user, what I really want is for this tool to automatically check and correct AI output before every response — a seamless experience with no manual steps.

I noticed that docs/mcp.md already has a prompt that achieves this, and src/mcp/setup.rs has subcommands that generate platform-specific instructions files. However, these all require users to discover and configure them on their own, and there is no clear guidance in the README.

Most users who rely on AI do not realize that the Chinese output of current AI models needs correction, and the AI itself does not know it is producing non-standard vocabulary either. If the MCP Server could work automatically after installation, it would greatly lower the barrier and reach the users who are unconsciously adopting incorrect terminology. This led me to research how MCP tools can be triggered automatically.

While looking into this, I found that Context7 recently conducted an invocation routing eval on exactly this question — whether AI would proactively call an installed MCP tool. The results showed:

• Context7's MCP server already returns an instructions field during initialization, telling the AI when to call it. Even so, the trigger rate still drops from 92% to 72% during ongoing conversations, because AI in working mode tends to answer from its own knowledge.
• Adding a second layer of instructions on top, such as rule files or CLAUDE.md, recovers the rate to 94-100%.

Whether this approach fits zhtw-mcp depends on the project's roadmap, but if the goal is to also reach users who do not realize their AI output needs correction, the instructions field in InitializeResult might be worth exploring.

[jserv]

Check the “Common prompts” section in README.md if you want to use zhtw-mcp. In daily use, I typically use the prompt: “use zhtw to refine .”

Practically, we can introduce Claude Code hook for automatic zh-TW linting.

• Implement PostToolUse hook that auto-invokes zhtw-mcp lint on
  files matching *.md, *.txt, *.yml when the tool is Write
  or Edit.
  • No bash script. Add zhtw-mcp internal hook-callback hidden
    subcommand that reads the Claude Code hook JSON payload from stdin,
    extracts file_path, checks extension, runs the lint, and writes
    the result JSON to stdout. All logic in Rust -- cross-platform.
  • Add zhtw-mcp hook install subcommand that writes the hook entry
    to ~/.claude/settings.json.
  • Token budget protection:
    1. Emit nothing (exit 0, empty stdout) when no issues found.
    2. Cache (path, blake3(content)) -> issue_set on disk. Reuse
       existing blake3 crate. Suppress output when issue set is
       unchanged from the previous invocation on same file content.
    3. Cap output to summary line + first 3 issues + "+N more".
  • Gate: hook adds <50ms to file-write operations. Zero tokens
    emitted on clean files.
  • Scope: src/main.rs (hook install + internal hook-callback
    subcommands).

In addition, Claude Code SKILL.md nd Gemini CLI skill generators with progressive disclosure (metadata loads first, full instructions activate only when relevant).