🤖 Claude generated, human verified

AI Attribution Scale — used throughout this repo:

A personal config kit for Claude Code: researched settings, reusable agents and skills, and meta-skills to keep it from going stale.

Figuring out the best Claude Code setup takes real research, and that research has a shelf life. This repo is where it lives, along with the config it produced, so nothing has to be re-derived from scratch.

• research/ Snapshots on the AI model landscape, prompting patterns, Claude Code setup, hooks, skill authoring, and usage limits. Each file is dated. Check the "Last updated" line before relying on specifics.
• tools/claude-code/ The deployable starting point: settings.json, CLAUDE.md, a statusline script, three hooks, seven reviewer subagents, and five skills. Nothing in here is project-specific. Once deployed, it's available across all your projects.
• .claude/skills/ Three skills for maintaining this repo itself (see below).

Skills

Hooks

The first two are registered in user-settings.json because nearly everyone wants them. writing-voice-guard.sh is not, because it blocks writes over a style opinion rather than a safety issue. It also only ever blocks newly introduced dashes, so editing a legacy file full of them is fine.

To turn it on, add this to the PreToolUse array in your ~/.claude/settings.json:

{
  "matcher": "Edit|Write|MultiEdit|Bash",
  "hooks": [
    { "type": "command", "command": "~/.claude/hooks/writing-voice-guard.sh" }
  ]
}

To turn it off again without unregistering it, set "WRITING_VOICE_GUARD_OFF": "1" in the env block of the same file. Note that an export inside a session does not reach it, since hooks are spawned from Claude Code's own environment.

To allow dashes only under specific paths, such as a raw cache of externally sourced content that must stay byte-for-byte faithful, set "WRITING_VOICE_GUARD_ALLOW_PATHS" to a colon-separated list of glob patterns instead, in a project's own .claude/settings.json so the exemption stays scoped to that project (e.g. "*/pulled/raw/*").

The banned characters are hardcoded, but the pattern generalizes. Any rule you can detect mechanically can work the same way: block on the detectable violation, then use the block message to ask for a full style pass rather than a one-character fix. research/prose-linting.md covers the reasoning and the tradeoffs.

Subagents: staff engineer, security, product, design, test quality, Cloudflare specialist, generic domain expert.

These run inside this repo only, not as part of your global setup.

• update-research Re-checks research against current sources. Run this periodically; snapshots drift.
• update-templates Brings tools/claude-code/ in line with the research. Run after update-research, or any time you learn something worth baking into the starting point.
• deploy-config Copies tools/claude-code/ to ~/.claude/. Safe to run more than once: missing files get created, identical files are skipped, and anything that already exists and differs gets flagged for your decision.

1. Clone the repo and open it in Claude Code.
2. Look through tools/claude-code/ and cut what does not fit your workflow.
3. Ask Claude to run the deploy-config skill to install everything to ~/.claude/.
4. Later, run update-research then update-templates to stay current.

• awesome-claude-code Curated directory of Claude Code hooks, MCP servers, slash commands, and agent collections.
• my-claude-code-setup A multi-file CLAUDE.md "memory bank" pattern worth a look if a single CLAUDE.md is not enough for your project.

Note: there is a CLI tool called claude-code-templates by davila7. Different project; the name overlaps with this repo's tools/claude-code/ folder.