Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

🏠 Back to Blog

CC4E #5: The Best Personal Assistant Remembers Things About You

Author: Hannah Stulberg | Published: 2026-02-14

A deep dive on CLAUDE.md files – onboarding documents that auto-load so Claude starts every session already briefed. Solves the “Groundhog Day employee” problem.

How CLAUDE.md Files Work

  • Plain text markdown files, named exactly CLAUDE.md (all caps required)
  • Auto-load at session start from the current folder all the way up to the root
  • Hierarchical loading: deeper files layer more specific context on top of general context
  • Conflict resolution: more specific (deeper) file wins
  • Applies equally to AGENTS.md (Cursor, Antigravity) – concepts are tool-agnostic

Critical Trade-Off

Every line in a CLAUDE.md file consumes thinking room from the start. More context loaded = less room to think = lower quality output. Be intentional about what goes in.

  • If CLAUDE.md files push context to 10-15% before you start, they’re too heavy
  • Practical ceiling: ~150 total instructions across all loaded files (Claude Code’s system prompt uses ~50, leaving ~100-150)

Folder Structure Controls Context

LevelLocationContains
Level 0~/CLAUDE.mdPersonal working style, portable across jobs
Level 1~/Acme/CLAUDE.mdCompany context, role, team, areas
Level 2~/Acme/Marketing/CLAUDE.mdDomain-specific (brand voice, processes)
Level 3~/Acme/Marketing/Q1Campaign/CLAUDE.mdProject goals, timeline, decisions, status

Only create a CLAUDE.md when you have context that is (1) specific to that level and (2) needed across multiple sessions.

Four Types of Content

  1. Document index – map of what’s where so Claude doesn’t search blindly. Can use @ imports for auto-loading specific files
  2. The people involved – names, roles, relationships at the relevant level
  3. What this is about – identity, goals, key facts
  4. How you want things done – workflows, processes, preferences, guardrails

What Does NOT Belong

  • Passwords, API keys, sensitive data
  • Context that belongs at a different level
  • Info that changes so frequently it’ll be stale next session
  • Templates (use custom slash commands instead – loaded on demand)
  • Litmus test: if a piece of context isn’t relevant to the vast majority of sessions where that file loads, it doesn’t belong

Three Methods to Write Them

MethodWhen to useHow
Plan mode interviewContext lives in your headPlan mode, Claude interviews you, you answer conversationally
Load docs, then generateYou have existing documentsSave docs into the folder, ask Claude to read them and draft
Generate from a sessionJust finished productive workAsk Claude to capture durable context from the session

Claude drafts, you edit. Always review – Claude may include session-specific details that aren’t durable.

Maintenance

  • After big sessions: ask Claude if CLAUDE.md needs updating
  • When you keep re-explaining the same thing: make it permanent
  • Signs a file needs trimming: longer than when created, finished project details, multiple projects crammed in one file
  • CLAUDE.md files load at session start. Mid-session edits require re-reading the file, /clear, or a new session