claudette

You've spent hours building a knowledge base of hard-won debugging insights, custom skills, and specialized agents. But Claude Code doesn't know they exist unless you remember to mention them — so you keep re-discovering the same race condition fix, re-explaining the same API quirk, losing context you already captured.

Claudette fixes that. Two commands, zero maintenance:

go install github.com/dotcommander/claudette/cmd/claudette@latest
claudette install

claudette install adds two hook entries to ~/.claude/settings.json (UserPromptSubmit and PostToolUseFailure), writes ~/.config/claudette/config.json, and builds the initial index. It's idempotent. Reverse it any time with claudette uninstall.

Now when you type "fix the goroutine deadlock" at 11pm, Claude automatically sees the KB entry you wrote three weeks ago — the one where you spent an hour tracing that channel bug. When your build breaks, Claude surfaces the entry from last time you hit that exact error. Your past debugging sessions become automatic context for every future conversation.

What changes

Before: You write fix goroutine race condition. Claude starts from scratch. Your KB entry about Go race patterns sits unread in ~/.claude/kb/go/.

After: Claudette scores your prompt against every KB article, skill, agent, and command you've installed. Claude sees the match, reads the entry, and applies what you already know — before writing a single line of code.

This works for errors too. A test fails with undefined: NewRouter — Claude Code fires PostToolUseFailure, claudette tokenizes the failing tool output, finds your KB entry about chi/v5 import paths, and surfaces it. You stop re-debugging solved problems.

How it works

Claudette runs as a Claude Code hook — invisible infrastructure that fires on every prompt:

1. UserPromptSubmit — scores your prompt against indexed entries and surfaces the top matches. Runs in under 50ms.
2. PostToolUseFailure — fires only when a tool call fails. Tokenizes the failing output and surfaces relevant KB entries so Claude sees the fix you already wrote down.

claudette install wires both hooks into ~/.claude/settings.json and builds the index. It's idempotent — safe to re-run anytime. claudette uninstall removes every entry it added and deletes ~/.config/claudette/; the binary itself stays in $GOPATH/bin so you can remove it with rm when you're ready.

CLI

You can also search your knowledge base directly:

claudette search goroutine patterns     # search all entry types
claudette kb sqlite connection pool     # KB entries only
claudette skill refactoring             # skills only
claudette scan                          # rebuild the index

Flags: --format json, --threshold 3, --limit 10

What gets indexed

Everything under ~/.claude/:

Directory	What's there
kb/	Knowledge base articles — debugging insights, API quirks, patterns
skills/	Skills — specialized capabilities and domain knowledge
agents/	Agent definitions
commands/	Slash commands
Plugins	Anything installed via ~/.claude/plugins/

The index lives at ~/.config/claudette/index.json and auto-rebuilds when files change. No manual maintenance.

Configuration

claudette install writes ~/.config/claudette/config.json:

{
  "source_dirs": [
    "/home/you/.claude/kb",
    "/home/you/.claude/skills",
    "/home/you/.claude/agents",
    "/home/you/.claude/commands"
  ]
}

Add extra directories to index team-wide skill repos or project-specific knowledge. Plugin directories are included automatically.

Variable	Values	Description
CLAUDETTE_OUTPUT	full (default), compact	Controls how much detail appears in surfaced entries

Documentation

• Installation — install, verify, uninstall, update
• How It Works — hook flow, index, scoring algorithm
• Searching — CLI commands and flags
• Configuration — source directories, output modes, env vars
• Writing Entries — frontmatter, keyword weights, testing your entries
• Troubleshooting — diagnostics, common errors, hook debugging
• Contributing
• Changelog

License

MIT