Every Claude Code session starts with an empty memory of your files. Claude does not reread your whole project at the start. It loads your CLAUDE.md, then searches and opens files when a task needs them. That search is what you pay for, again in every new session.
A map is one short file that says what lives where. Claude reads it at the start and goes to the right file faster.
My test (11 Oct 2026): an open-source online store (78 files), 3 everyday tasks, each run twice with and without a map. Same answers, all correct. With the map: 24% fewer tool calls, 21% fewer input tokens processed, 9% faster. Best task: 1 search instead of 2, 34% fewer tokens. The dollar cost only dropped 2 to 10%, because Claude Code already caches repeated context cheaply. No 70% here.
Save the map skill (SKILL.md in this gist) as:
~/.claude/skills/map/SKILL.md
Use .claude/skills/map/SKILL.md inside a project instead if only that project should have it.
Open Claude Code in your project folder and type:
/map
Claude writes MAP.md (40 to 80 lines: what the project is, folders, key files, where to change things,
which settings exist) and adds this line to your CLAUDE.md:
Project map: @MAP.md. Check it before searching. If it is wrong, fix MAP.md in the same task.
The @ makes Claude Code load the map at the start of every session.
Start a new session and run /context. Your CLAUDE.md should be listed under Memory files.
Then ask something like "where do I change X?" and watch: Claude should open the right file with fewer searches.
Run /map again after you add, move or delete folders or important files. It updates the map instead of rewriting it.
- Small projects don't need it. Under ~30 files, Claude finds things fast anyway.
- It costs a little every session. My map was about 1,900 tokens, loaded each time. It pays off when tasks would otherwise need several searches.
- A wrong map is worse than none. That's why the CLAUDE.md line tells Claude to fix it when it's off.
- Making the map costs once. About $0.27 and 24 seconds in my test.
- Bigger wins usually come from elsewhere: specific prompts ("fix the bug in auth.ts"),
/clearbetween unrelated tasks, and fewer always-loaded skills and MCP servers. In my setup, ~91,000 tokens were loaded before Claude opened a single file.
Sources: code.claude.com/docs/en/memory (what loads at start, @ imports), /context-window (what loads before you type),
/costs (a project-overview skill saves exploring). Numbers are from one project on one machine: mine.
— Conrad · @buildwith.conrad