Skip to content

Auto-Context — auto-context.py

Runs on every UserPromptSubmit. Reads the prompt from stdin JSON, matches against a regex/project table, emits a small markdown pointer-list for the matched projects. Silent when nothing matches.

How it decides

Keyword → project mapping is defined in the script itself (PROJECTS list). Each entry: - regex pattern (case-insensitive, word-boundary anchored) - project label - list of file/dir pointers

Top 3 matches surface. Skipped on prompts shorter than 4 chars.

Current project coverage

Nova Design, JustinWieb-VR, Home Assistant, Frigate/Cameras, Business Second Brain, justinandkrystal.com, Plex/Media, n8n, Greg/OpenClaw, Google Drive, Wiebelhaus Enterprises, Web Builder mode, Context API, Wellness pollers, Telegram bot, Forge fleet platform.

Edit when

  • A new project is added — append a regex+label+pointers entry to PROJECTS in scripts/hooks/auto-context.py
  • A project is renamed or moved
  • You notice the hook firing on false-positive keywords (tighten the regex)

Design rules

  • Never crashes (wrapped in try/except, swallows all exceptions)
  • Silent on non-project prompts (most prompts)
  • Emits pointers not contents — Claude decides which to Read
  • Caps at 3 matches per prompt to avoid context bloat

Files

  • scripts/hooks/auto-context.py — the hook
  • .claude/settings.local.json → hooks.UserPromptSubmit wires it up (chained after prompt-counter.sh)