TomeVault/instruction-corpus

The TomeVault Instruction Corpus

De-identified structural measurements of AI instruction files (CLAUDE.md, AGENTS.md, SKILL.md, .cursorrules and related conventions).

Edition 2026-07. 229,720 files. Schema version 1.0.0. Licensed CC BY 4.0.

What this is

Teams increasingly commit instructions for AI coding agents directly into their repositories. CLAUDE.md, AGENTS.md, SKILL.md, .cursorrules and a dozen sibling conventions are now ordinary source-controlled artefacts, and almost nothing is measured about them.

This dataset is one row per instruction file that our pipeline has fetched and parsed, carrying that file's structural properties. Size, format, licence, whether it names a model that no longer exists, and whether a deterministic ruleset thinks an agent asked to load it would get what the author intended.

It contains no file contents, no owner names, no repository names and no URLs. Every measurement in it is reproducible from the method described below.

One row per instruction FILE INSTANCE the pipeline has fetched and converted. A row is a file, NOT a repository and NOT an owner: one repository commonly contributes many rows, and a widely-forked file appears once per fork. Any count derived from this table is a count of files unless it explicitly de-duplicates. Files whose deterministic security grade is 'fail' are excluded. Each file is measured once, when first seen, and the measurements are retained permanently, so a file stays in the population for every later edition even after we discard its body for storage reasons. Files first crawled before that retention policy began are absent, which makes early editions a floor rather than a census.

Source and updates

This is one edition of a monthly series. Every edition is dated, immutable once published, and keeps its own citable identity, so a figure you quote does not change underneath you.

• Series index and past editions: https://tomevault.io/standards/state-reports
• Method, column definitions and the checks behind them: https://tomevault.io/standards/state-reports
• Published by TomeVault: https://tomevault.io
• Questions, corrections, or removal requests: oli@tomevault.io

If you use this dataset in something public we would genuinely like to see it.

What this cannot tell you

Three limits are worth stating before anyone builds on this.

A row is a file, not a project and not a team. A popular instruction file that has been forked four hundred times contributes four hundred rows. Any count here is a count of file instances unless you de-duplicate it yourself, and the difference between the two is large.

The observation window is ours, not the ecosystem's. The ingested_week column records when our crawler first saw a file, which is not when it was written. Growth curves drawn from that column describe our crawl, not adoption.

Structural checks are not quality judgements. A file can pass every check here and still give an agent bad instructions. What the loadability columns measure is whether the instruction reaches the agent intact, which is a narrower and more testable question than whether it is any good.

Columns

Column	Type	Meaning
format	string	The instruction-file convention this file follows, as detected by the pipeline (claude_md, agents_md, skill_md, cursorrules, cursor_mdc, gemini_md, copilot_instructions, windsurf_rules, other).
format_family	string	The format grouped to its tool family (claude, agents, cursor, gemini, copilot, windsurf, skill, other). Formats that a single tool reads under more than one filename collapse to one family here.
role	string	Whether the file is an always-on config, an on-demand skill, or a tome.
source_platform	string	The platform the file was discovered on.
account_type	string	Whether the owning account is an Organization or an individual User, as reported by the source platform. Null where the platform did not say.
bytes	int64	Size of the file body in bytes.
lines	int32	Number of newline-separated lines in the file body.
token_count	int32	Approximate token count of the body, as computed by the pipeline at ingestion. Null where it was never computed.
licence_spdx	string	SPDX identifier of the REPOSITORY's licence as reported by the source platform. Null means the platform reported none, which is not the same as the work being public domain. This population is licence-filtered, so the distribution here is not the ecosystem's: see the licence note in the biases section before deriving any rate from it.
names_a_model	bool	True when the body names at least one model identifier from the curated tracking list, at any lifecycle status.
model_is_retired	bool	True when the body names at least one model identifier whose curated status is deprecated, superseded or retired. False when the file names only live models AND when it names no model at all. Read this column together with names_a_model: the share of the whole corpus and the share of model-naming files are very different numbers, and only the first is a statement about the corpus.
retired_model_count	int32	How many DISTINCT retired identifiers the body names.
loadability_tier	string	The verdict of the deterministic loadability ruleset over this file: whether an agent asked to load it would get what the author intended.
loadability_findings	int32	Number of distinct loadability checks that fired at least once on this file. Zero means the file passed every applicable check.
has_broken_reference	bool	True when at least one loadability finding is a reference the file makes to something that will not travel with it, or that is already gone.
security_grade	string	Grade from the deterministic pattern scanner. Rows graded 'fail' are not in this dataset at all; this column distinguishes the rest.
quality_grade	string	Grade from the deterministic quality ruleset. Null where never graded.
ingested_week	string	ISO week (YYYY-Www) in which the pipeline first observed the file. See the ingestion-window caveat in the dataset card.

Method

Files are discovered by crawling public repositories, fetched, and classified by format. Every measurement in this release comes from deterministic rulesets, not from a language model, so the same input always produces the same row.

Loadability findings come from the same ruleset our product runs, version None. The has_broken_reference column is true when a file fires either the stale-path check or the reference-portability check, meaning it points at something that has gone or that will not travel with the file when it moves.

Retired-model flags come from a hand-curated list of model identifiers and their lifecycle status, last updated 2026-07-31. Only identifiers marked deprecated, superseded or retired are flagged. Matching is whole-token, so gpt-4 does not match inside gpt-4o.

Licence classification reuses the same policy module that gates our own ingestion, so the compliance figures here and our enforcement are definitionally the same thing.

aggregate.json carries the headline metrics from snapshot 2026-07, which is the measurement run this edition is built from. The two files answer different questions and do not share a denominator: the parquet is one row per retained-body file instance, while each metric in the aggregate declares its own population in its own population field. Read a figure with the population beside it, and do not reconcile a count from one against a count from the other.

Known biases

Read model_is_retired together with names_a_model. Most files name no model at all, and a file that names no model cannot be stale. The share of the whole corpus naming a retired model and the share of model-naming files naming a retired model are very different numbers. Only the first is a statement about the corpus. Publishing the second as though it were the first overstates the problem by more than an order of magnitude, and we would rather you did not.

The curated model list carries more dead identifiers than live ones. That shape inflates any conditional staleness rate independently of how careless anyone has been.

Discovery is not uniform. Coverage reflects what a crawler over public repositories reaches. Private repositories, self-hosted platforms and files inside archives are absent, and there is no reason to assume the absent population resembles this one.

Early editions are a floor, not a census. Each file is measured once, when first seen, and those measurements are kept permanently. Files first crawled before that retention began were never measured and cannot be recovered without re-fetching them, so they are missing. Counts will therefore rise across the first few editions partly because coverage is filling in, not only because the ecosystem is growing. Treat early growth in absolute counts with suspicion and prefer the proportions, which are far less affected.

Security-failing files are excluded. Rows whose deterministic security grade is fail were removed before export. The remaining population is therefore slightly cleaner than the raw corpus.

The licence distribution is filtered, and it is not the ecosystem's. Files whose repository licence fails our ingestion policy do not reach this population at all, so permissive licences are heavily over-represented here by construction. A small share of rows additionally carry no SPDX identifier, because the source platform reported none. Neither the shape of this column nor the size of its null slice supports any claim about how much of the wider ecosystem is unlicensed or copyleft. Do not read a compliance rate out of this dataset.

ingested_week is the ISO week in which OUR PIPELINE first observed the file, not the week the file was authored upstream. The series describes the pipeline's observation window and must never be presented as ecosystem growth. Deriving an adoption curve from this column is a misuse of it.

Removal and consent

This release excludes every file belonging to an owner who has asked to be removed. Owners are excluded at the OWNER level. The exporter first collects every owner having at least one row with opted_out = 1, then drops all rows belonging to those owners. Row-level filtering would leak the untouched rows of an owner who has already asked to be removed.

Edition 2026-07 excludes 275 files across 16 owners on that basis.

To be removed from future editions, contact oli@tomevault.io. Because the dataset carries no owner names or repository names, removal takes effect by dropping the underlying rows, and subsequent editions simply do not contain them.

Cite this

Plain text:

TomeVault (2026). The TomeVault Instruction Corpus, edition 2026-07. https://tomevault.io/standards/state-reports

BibTeX:

@dataset{tomevault_instruction_corpus_2026_07,
  title  = {The TomeVault Instruction Corpus},
  author = {TomeVault},
  year   = {2026},
  note   = {Edition 2026-07, schema 1.0.0},
  url    = {https://tomevault.io/standards/state-reports}
}

Corrections ship as new dated editions. A published edition is never edited in place, so a figure you quote does not change underneath you.