README.md

---
language:
  - en
size_categories:
  - 1M<n<10M
task_categories:
  - text-generation
tags:
  - code
  - agents
  - agents-md
  - cursor-rules
  - claude-skills
  - github
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*.parquet
---

# rules

Natural-language LLM-agent rule files (`AGENTS.md`, `CLAUDE.md`, `SKILL.md`, `.cursor/rules/*.mdc`,
and friends) crawled from public GitHub repositories, with the content stored inline.

**2,204,470 files from 45,014 repositories.** Each row is one rule file, pinned to the commit SHA
it was read at, so `link` always resolves to the exact bytes in `file`.

## Columns

| column | type | description |
| --- | --- | --- |
| `file` | string | full text of the rule file |
| `content_sha256` | string | SHA-256 of `file` |
| `path` | string | path within the source repository |
| `link` | string | permalink to the file at the pinned commit |
| `repo` | string | `owner/name` of the source repository |
| `stars` | int64 | stargazer count of the source repository at crawl time |
| `crawled_time` | timestamp | when the file was fetched |

## Composition

No single filename dominates — `AGENTS.md` is under 3% of rows. Skill bundles contribute the most
files, while the named manifests reach the most repositories.

| family | rows | share | repos |
| --- | ---: | ---: | ---: |
| `SKILL.md` (Claude skills) | 1,088,809 | 49.4% | 15,835 |
| supporting `.md` inside `.claude/`, `.cursor/`, … | 655,713 | 29.7% | 19,415 |
| `agents/*.md` (subagent definitions) | 139,997 | 6.4% | 8,592 |
| `.mdc` (Cursor rules) | 77,911 | 3.5% | 8,022 |
| `AGENTS.md` | 60,466 | 2.7% | 15,924 |
| `CLAUDE.md` | 52,717 | 2.4% | 18,869 |
| `.prompt.md` | 48,898 | 2.2% | 4,738 |
| `.instructions.md` | 34,364 | 1.6% | 4,951 |
| `.cursorrules` | 13,234 | 0.6% | 5,278 |
| `GEMINI.md` | 10,754 | 0.5% | 3,946 |
| `copilot-instructions.md` | 9,476 | 0.4% | 8,618 |
| `.chatmode.md` (Copilot) | 7,699 | 0.3% | 1,801 |
| `.windsurfrules` | 4,053 | 0.2% | 3,389 |
| `.goosehints` | 379 | 0.02% | 323 |

## Usage

```python
from datasets import load_dataset

ds = load_dataset("Andwwy/rules", split="train")
```

Or query the parquet directly, without downloading, from DuckDB:

```sql
INSTALL httpfs; LOAD httpfs;

-- ALL rule files, most-starred first
SELECT repo, path, stars
FROM read_parquet('hf://datasets/Andwwy/rules/data/*.parquet')
ORDER BY stars DESC
LIMIT 20;
```

See what the corpus is made of:

```sql
SELECT CASE
         WHEN path ILIKE '%SKILL.md'                THEN 'SKILL.md'
         WHEN path ILIKE '%.mdc'                    THEN 'cursor .mdc'
         WHEN path ILIKE '%AGENTS.md'               THEN 'AGENTS.md'
         WHEN path ILIKE '%CLAUDE.md'               THEN 'CLAUDE.md'
         WHEN path ILIKE '%.prompt.md'              THEN '.prompt.md'
         WHEN path ILIKE '%.instructions.md'        THEN '.instructions.md'
         WHEN path ILIKE '%.chatmode.md'            THEN '.chatmode.md'
         WHEN path ILIKE '%.cursorrules%'           THEN '.cursorrules'
         WHEN path ILIKE '%.windsurfrules%'         THEN '.windsurfrules'
         WHEN path ILIKE '%.goosehints%'            THEN '.goosehints'
         WHEN path ILIKE '%GEMINI.md'               THEN 'GEMINI.md'
         WHEN path ILIKE '%copilot-instructions.md' THEN 'copilot-instructions.md'
         WHEN path ILIKE '%/agents/%'               THEN 'agents/*.md'
         ELSE 'other'
       END                  AS family,
       count(*)             AS files,
       count(DISTINCT repo) AS repos
FROM read_parquet('hf://datasets/Andwwy/rules/data/*.parquet')
GROUP BY family
ORDER BY files DESC;
```

Narrow to one ecosystem — everything Cursor reads, say:

```sql
SELECT repo, path, stars, file
FROM read_parquet('hf://datasets/Andwwy/rules/data/*.parquet')
WHERE path ILIKE '%.mdc' OR path ILIKE '%.cursorrules%'
ORDER BY stars DESC;
```

Or pull the whole rule set for a single repository:

```sql
SELECT path, length(file) AS bytes, link
FROM read_parquet('hf://datasets/Andwwy/rules/data/*.parquet')
WHERE repo = 'anthropics/claude-code'
ORDER BY path;
```

## Collection

Repositories are discovered via the GitHub Code Search API, pinned to a commit via GraphQL,
enumerated with the Git Trees API, and the file bodies are fetched from `raw.githubusercontent.com`
at that commit. Code search selects *repositories* only — it caps at 1,000 results per query and
surfaces only a fraction of a repo's matching files — so the file list always comes from the Trees
API, which returns an entire repository in one request.

Deduplicated on `(repo, path, commit)`. Identical text appearing in different repositories is kept
as separate rows: 2,200,687 distinct files span 1,088,448 distinct contents, and that redundancy is
itself signal about what gets copied. Symbolic links, and files whose entire body is a pointer to
another file (e.g. `@AGENTS.md`), are dropped.

## Licensing

Each file remains under the license of the repository it came from; `link` and `repo` identify
the source. No single license applies to the collection as a whole.