AutoADHD: Instructions file for Claude Code

CLAUDE.md

You are an AI agent operating inside an Obsidian vault. Git tracks every change. If something goes wrong, we revert.

---

WHAT YOU CAN DO

• Create, edit, and delete notes in any folder
• Add new Canon entries directly (people, events, concepts, decisions, projects, actions — or new types as needed)
• Update existing Canon entries with new information
• Create and maintain wikilinks between notes
• Evolve the structure: new types, new relationship patterns, new folders are fine
• Research external context for concepts/ideas and add with proper source attribution
• Write AI reflections and observations to Meta/AI-Reflections/

WHAT YOU MUST NOT DO

• Never edit content under # Raw in any Inbox note. Raw transcripts are source evidence and are read-only.
• Never delete Inbox notes that contain raw transcripts
• Never fabricate information. If you're unsure, say so.

---

SKILLS (just-in-time context)

Skills inject focused rules at the moment you need them, instead of relying on this file surviving a long conversation. They live in .claude/skills/ and are loaded on demand.

Skill	When to load	What it contains
vault-writer	Creating or editing ANY vault note	Note format, frontmatter, emoji headings, provenance, wikilinks, folder map
vault-extractor	Processing an Inbox note into Canon entries	Full extraction checklist, action detection, reflection handling, duplicate detection, write-back rules

For Cowork sessions: Skills load automatically when triggered by the task description.

For CLI/Codex agents: Agent scripts should tell the agent to read the relevant skill file at runtime (lean prompt pattern) rather than dumping rules into the CLI argument. Example: "Read .claude/skills/vault-writer/SKILL.md before creating any notes."

Token discipline: Skills replace the need to load this entire CLAUDE.md into every agent prompt. Load only what the task requires. The vault-writer skill (~200 lines) replaces the note format, emoji, provenance, and wikilink sections of this file (~150 lines) — but with fresher context at the point of action.