An instructions file is read whole, at the start of every session, by every agent that supports it. You pay for it whether or not the session turns out to need it. The median file in the catalogue is 1,148 tokens per session, paid before you ask your first question.
What your agent reads before you type
Each agent reads the filename it knows. AGENTS.md is read by Codex, Cursor, Copilot’s agent mode and most others: 14,769 files, median 1,210 tokens. CLAUDE.md is read by Claude Code: 14,484 files, median 1,296. copilot-instructions.md by GitHub Copilot: 2,055 files, median 685. GEMINI.md by Gemini CLI: 758 files, median 512.
So a repository with both an AGENTS.md and a CLAUDE.md pays for one of them per session, not both. What it pays twice for is keeping them in step. Claude Code does not read AGENTS.md at all. The usual bridge is a one-line CLAUDE.md that imports it with @AGENTS.md, which loads the same text at the same cost.
The 1,148-token default
A token is the unit your AI plan bills by, roughly three quarters of a word. The median instructions file costs 1,148 of them every session. Twenty sessions a day make that 22,960 tokens before any work starts. The 90th percentile is 4,611. The largest in the catalogue is FCoP, 55,112 tokens across 3,703 lines, every session.
The Next.js AGENTS.md is 7,296 tokens across 561 lines. The Supabase AGENTS.md describes the same kind of monorepo in 76 lines and 1,370 tokens. The difference is what each decided the model must know before every session, as opposed to what it can be told when it asks.
What instructions files tell agents to do
31,679 of the 34,245 files are graded A, 1,182 B, 1,372 C or D and 12 E or F. An instructions file is plain text, so a finding is about what the text tells the agent to do. The most common: 2,134 files tell the agent to make network calls, 642 point it at the agent’s own configuration directories, 617 contain a recursive force delete.
Most of those are ordinary in context, a curl in a setup step or an rm -rf in a clean script. Each finding is printed with the line of the file that produced it, so you can read it and decide. The OpenHands AGENTS.md is graded E for three findings at once: it downloads and runs remote code, makes network calls and force-deletes. Read the lines before you decide what that means for you.
The larger risk is drift. 5,340 files changed in the last seven days. An instructions file is edited like source code, and unlike source code the change lands in every future session without a review. The file’s page shows what changed since the last crawl.
How large projects keep theirs
- Next.js: the monorepo layout and how to work in it. 7,296 tokens every session, 561 lines.
- Codex: OpenAI’s own instructions for the Codex codebase, Rust included. 5,153 tokens every session, 321 lines, 6 copies.
- Spec Kit: the project instructions for a toolkit that has teams write the specification first. 7,104 tokens every session, 602 lines.
- LangChain: the packages and the development commands. 4,469 tokens every session, 386 lines, 1 copy.
- DeepSeek Harness: layout, rules and the pre-release approach. 3,735 tokens every session, 156 lines, 7 copies.
- Supabase: structure, tools and conventions for a monorepo. 1,370 tokens every session, 76 lines.
What belongs somewhere else
When the text is needed rarely. An instructions file cannot be conditional: it loads whether the session is about the build system or a typo in the README. Text that matters one session in many belongs in a skill, a pack of instructions the model opens only when it needs them, 68 tokens until then. Text about one file type belongs in a rule with a glob, a file pattern, so it loads only when a matching file is open. And when you need something to run rather than to be read, that is a hook, not a line in a file.
Questions people ask
Does Claude Code read AGENTS.md? No. It reads CLAUDE.md. A one-line CLAUDE.md can import AGENTS.md with @AGENTS.md, and then the same text costs the same every session.
Is one big file or several small ones cheaper? Neither, if all of them load. A session pays for the sum of what is in its context, however that text is split, and all 34,245 files on this shelf are read whole. Splitting only saves anything if some of the pieces then stay out of the session.
How do I know what mine costs? npx agentmods cost prints what the mods in your project cost per session. Compare yours with the median, 1,148, and the 90th percentile, 4,611. The shelf spans 26,742 repositories, so there is plenty to compare against.
Do it now
Open your repository’s instructions file on this site: cost per session, what changed since the last crawl, any finding with its line. If yours is past 4,611 tokens, the section you need least often is the first candidate to move into a skill.