CLAUDE.md

The root CLAUDE.md is the index Claude reads first when a session opens. Its job is to point at the other files and set the universal rules — not to contain detail.

The most common mistake is letting this file grow. Every time you discover a new rule, the urge is to add it here. Resist. The right home is almost always a context file or a workflow file.

What to include

• Your role in two or three lines so Claude has a coarse sense of who you are
• A list of the files in context/ with one-line descriptions of each
• Three to five default behaviours that apply to every task regardless of workflow
• The set of trigger phrases that map to specific workflows

What to exclude

• Any actual content from the other files (no priority rules listed here, no people listed here)
• Code or detailed instructions for specific workflows
• Project-specific context — that lives in context/projects.md
• Voice examples — those live in context/voice.md

Length target

Under 80 lines. If CLAUDE.md grows past that, the model starts ignoring parts of it. The file is most effective when it’s short enough that every line earns its place.

Template

# Command Centre — [Your Name]

## Role
[One or two sentences on what you do, who you work for, and the
nature of your work. Enough that Claude knows the register and
audience for everything generated.]

## How to use this folder
Read only what is relevant to the current task. Each file in
context/ has a specific purpose — pull in the ones that apply,
ignore the rest.

## Context files (read on demand)
- context/role.md — what I do, who I report to, what I'm
  measured on
- context/priorities.md — how to triage what matters in my
  inbox and calendar
- context/people.md — clients, internal stakeholders, recurring
  contacts
- context/projects.md — active workstreams and their state
- context/voice.md — my writing style with real examples
- context/anti-voice.md — words, patterns, and structures Claude
  must never use in anything written for me

## Default behaviours
1. Never send emails. Always produce drafts I review.
2. For PowerPoint output, use templates/brand_template.pptx as
   the base unless I say otherwise.
3. Save all generated decks to outputs/decks/ with date prefix
   YYYY-MM-DD.
4. When uncertain about facts, names, or numbers, flag the
   uncertainty rather than guess.
5. Drafts go to outputs/drafts/ as .eml or .md files.

## Common commands
- "Triage my inbox" → use workflows/triage.md
- "Run the morning brief" → use workflows/brief.md
- "Draft a reply to [thread]" → use workflows/reply.md
- "Build a deck from [outline]" → use workflows/deck.md

Why this file is short by design

Claude Code injects a system reminder telling the model that CLAUDE.md context may not be relevant to the current task. The longer the file, the more aggressively the model discounts it. A 200-line CLAUDE.md with detailed rules across five different workflows is worse than a 40-line index that points at the workflow file for the current task.

Treat additions to CLAUDE.md the way you’d treat additions to a system prompt: high cost, only justified by genuinely universal rules.