Answers
Leave the directory tree out of CLAUDE.md and AGENTS.md: it goes stale and the agent can list files itself
Published
A directory tree pasted into CLAUDE.md or AGENTS.md is the first part of the file to go stale, and the agent does not need it: Claude Code and Codex can list the repository themselves. Replace the tree with one line per surprise, such as "webhooks live in payments/webhooks/, not api/", and delete it when the code moves.
For Claude Code and Codex. Vendor facts are from the pages linked below, read on November 18, 2026.
What to write instead
- Find trees: search the file for box-drawing characters such as
├──and└──. - Delete the tree. Keep only the paths an agent would guess wrong, as a sentence each.
- Point at the source of truth for structure instead, for example "packages are listed in pnpm-workspace.yaml".
- When a pull request moves a folder, search the rule files for the old path in the same PR.
grep -nE "├──|└──" CLAUDE.md AGENTS.mdThe numbers
| Measure | Value | Source |
|---|---|---|
| Public CLAUDE.md files that contain a directory tree | 16.9% (weighted, of 1,932) | Harbor study, 18 Oct 2026 |
| Public AGENTS.md files that contain a directory tree | 10.5% (weighted, of 1,966) | Harbor study, 18 Oct 2026 |
| CLAUDE.md files past Anthropic’s 200-line target | 18.9% (weighted) | Harbor study, 18 Oct 2026 |
| Median CLAUDE.md length | 81 lines; 1 in 10 longer than 309 | Harbor study, 18 Oct 2026 |
Study figures are weighted estimates from 1,935 CLAUDE.md and 1,970 AGENTS.md files sampled from public GitHub repositories on 18 October 2026, with GitHub's own counts for the totals. Every figure is in the study's CSV, and the method is in the study.
What the docs say
- Claude Code memory docs: Target under 200 lines per CLAUDE.md file. Longer files consume more context and reduce adherence.
- Codex: Custom instructions with AGENTS.md: All project AGENTS.md files share a 32 KiB budget by default, so lines that describe the layout are spent before the rules below them.
Where Harbor fits
Harbor connects Claude Code, Codex, Cursor and GitHub Copilot (the CLI and VS Code agent mode; the cloud agent is not supported yet) through hooks and an MCP server. It ties rules to the code they are about: when a push deletes the file a rule names, the rule is retired, or flagged if other repos still use it, instead of sitting in a file nobody re-reads. Each rule shows how often it was served and how often an answer cited it. A citation is evidence a rule was used, not proof it was obeyed.
Read next
- Context rot and stale rules
- How long should CLAUDE.md be?
- The study behind the numbers
- CLAUDE.md and AGENTS.md drift apart in 7% of repos that copy one into the other: import instead
- Claude Code strips HTML comments from CLAUDE.md, so a rule inside one never loads
- "Read AGENTS.md" in CLAUDE.md does not load it: put @AGENTS.md on its own line
- Codex reads 32 KiB of AGENTS.md and cuts the rest: raise project_doc_max_bytes or trim
- Every answer, by agent and problem
Questions
Should I put a directory tree in CLAUDE.md?
No. It costs lines on every turn, it goes stale the first time a folder moves, and Claude Code can list the files itself. Write only the paths an agent would get wrong.
How many CLAUDE.md files include a directory tree?
In our 18 October 2026 sample, 16.9 percent of public CLAUDE.md files and 10.5 percent of AGENTS.md files (weighted) had box-drawing characters on two or more lines.
How do I stop rule files going stale?
Keep only what the code cannot tell the agent, put each rule in the folder it is about, and change the rule file in the same pull request that changes the code. A check in CI that greps rule files for deleted paths catches the rest.