Report
State of agent instruction files
Loading the dataset…
Across editions
The headline numbers of every monthly edition. Each edition measures the most-starred active repositories at the time, so the set of repositories changes between editions, and each records the turntax version that counted it. The highlighted month is the edition shown on this page.
Per-turn instruction tokens
o200k tokens for the agent that loads the most; a symlinked file counted once.
Repositories with each file
Share of the repositories with at least one file of that name.
Values by edition
Instruction tokens sent with every turn
One bar per range of per-turn size. For each repository the size is that of the agent that loads the most always-on instruction text from the repository's files (o200k_base tokens).
Which files repositories use
Repositories with at least one file of each kind. A repository usually has several kinds.
What the checks found
The same static checks the turntax CLI and web page run. A finding is a pointer to look at, not a verdict: some are deliberate choices.
Largest per-turn instruction context
The ten repositories whose files add the most always-on instruction text for one agent. Large projects, monorepos and agent tools often carry more guidance on purpose; size alone says nothing about quality.
All repositories
Select a column header to sort. Each name opens the live turntax report for that repository, which reads its current default branch.
Methods and limitations
Selection
- GitHub repository search
, sorted by stars (descending), on . - Repositories were taken in that order. A repository counts when its file tree has at least one agent instruction file that turntax recognises, outside dependency and build folders and outside test-data folders ().
- Every file was read at the default branch's head commit, pinned by SHA (listed in the table and the dataset). Files that instruction files pull in with
@pathimports were read too. Symbolic links were followed one level.
Counting
- Which files each agent loads, and when, follows each vendor's documentation as encoded in turntax . The audit assumes the agent is started at the repository root.
- Per-turn tokens are the always-loaded repository files only. User-level files (for example
~/.claude/CLAUDE.md), MCP server tool definitions and each agent's own system prompt are not counted, so real per-turn context is larger. - Counts use OpenAI's o200k_base tokenizer as a common reference. They are not Claude or Gemini tokenizer counts, which differ. Percentiles interpolate linearly between the closest ranks.
- Numbers are per turn before prompt caching. With caching, most of this text is billed at the cache-read rate after the first turn, but it still takes up context window.
- Findings and fixes come from static analysis of the files alone (the same as
npx turntaxandturntax --fix). Nothing was run. In a local checkout the CLI also checksnpm runscript names against package.json, so it can report a few more missing references. - The live report behind each repository link reads the current default branch and includes test and example folders, so its numbers can differ from this snapshot.
Reproduce
A scheduled GitHub Actions workflow collects a new edition on the first day of each month. To reproduce this edition's selection, run npm ci && npm run build in a checkout of turntax, then . The dataset records the turntax version and the commit of the counting engine. Search ranking and repositories change over time; the pinned SHAs let you re-run turntax on exactly the files measured here.