A searchable collection of 100 real-world AGENTS.md examples for AI coding
agents. Browse by stack, compare testing and safety patterns, or generate a
source-backed AGENTS.md template. Every example points to a fixed Git commit
and stores an integrity fingerprint.
Open the live gallery · Browse 100 examples · Compare patterns · Build a template · Contribute a source · Read the methodology
The fastest path is the interactive gallery:
- Search for a language, repository, or rule topic.
- Open the fixed GitHub source and read it in repository context.
- Compare its rules with the eight repeated patterns.
- Use the starter builder only as a small base—not a universal answer.
The index currently covers TypeScript, Python, Go, Rust, C, C++, Java, JavaScript, Kotlin, Swift, and other project stacks.
You can also open dedicated, indexable directories for technology stacks and rule patterns.
AGENTS.md is easy to create and surprisingly hard to write well. Generic
templates usually produce generic rules. Large instruction dumps consume
context without improving decisions.
This project takes a stricter route:
- look at files that real open-source projects actually maintain;
- pin every observation to a source commit;
- separate recurring patterns from claims about effectiveness;
- help people build a small starting file they can adapt to their repository.
The index contains 100 verified sources, 8 descriptive rule patterns, and 0 invented examples.
This is not another directory of generic templates. Every card follows the same evidence chain:
repository → full commit SHA → fixed AGENTS.md → SHA-256 fingerprint
- The source file is fetched directly from GitHub before inclusion.
- Tags describe visible content; they do not pretend to measure effectiveness.
- A scheduled check detects when the upstream file changes after the snapshot.
- The starter builder is derived from repeated patterns but stays deliberately small enough to adapt.
Important
Repository stars are captured only to help discovery. They are not treated as
evidence that an AGENTS.md file is effective.
| Repository | Stack | What makes the file interesting | Source |
|---|---|---|---|
radareorg/radare2 |
C | Location map, formatting, memory safety, commands | fixed commit |
google/benchmark |
C++ | Compact policy for reviewing AI-assisted work | fixed commit |
GitoxideLabs/gitoxide |
Rust | Test-first development, errors, commits, paths | fixed commit |
replicate/cog |
Go | Detailed command index and language conventions | fixed commit |
loro-dev/loro |
Rust | Short project snapshot and context index | fixed commit |
nextdns/nextdns |
Go | Core principles, tests, dependencies, binary budget | fixed commit |
DataDog/stratus-red-team |
Go | A focused 21-line contribution guide | fixed commit |
kubernetes-sigs/cluster-api-provider-openstack |
Go | Legal requirements, PR labels, Make targets | fixed commit |
The interactive gallery includes all 100 sources, filters, source metadata, pattern counts, and a stack-aware starter builder.
The current taxonomy is deliberately small:
| Pattern | Question |
|---|---|
| Testing | Does the file name useful test commands or test scope? |
| Scope | Does it explain where rules apply or how nested context works? |
| Safety | Does it define destructive or sensitive boundaries? |
| Code style | Does it point to formatting or existing conventions? |
| Verification | Does it say what evidence is required before handoff? |
| Git workflow | Does it protect work or define commit/PR expectations? |
| Generated files | Does it identify generated artifacts and their owner? |
| Documentation | Does it point to deeper project context? |
These tags describe content. They do not prove that a rule improves an agent. See the methodology for inclusion and evidence rules.
The starter builder generates a compact starting file for TypeScript/Next.js, Python, Go, or Rust. It can include:
- destructive-action boundaries;
- generated-file ownership;
- Git and publishing boundaries;
- real install, test, and build command placeholders.
The output is intentionally short. Replace every placeholder with commands from your repository and delete any rule you cannot enforce.
The machine-readable dataset is available at
data/sources.json. Each record includes a full Git commit
URL, descriptive tags, a pinned-content SHA-256 fingerprint, and the upstream
fingerprint observed on the snapshot date.
npm run sources:verify
npm run sources:verify-remoteThe scheduled GitHub Actions workflow checks pinned files and detects later upstream changes.
.
├── app/
│ ├── catalog-data.ts # source manifest and descriptive tags
│ ├── catalog-client.tsx # search, filters, patterns, starter builder
│ ├── discovery-page.tsx # shared stack and pattern directory UI
│ ├── globals.css # responsive visual system
│ ├── patterns/ # indexable pattern directory and detail pages
│ ├── sitemap.ts # generated route inventory
│ ├── tech-stacks/ # indexable stack directory and detail pages
│ └── page.tsx
├── data/
│ └── sources.json # 100 fixed sources plus integrity fingerprints
├── docs/
│ └── METHODOLOGY.md
├── scripts/
│ ├── discover-sources.mjs
│ └── verify-sources.mjs
├── tests/
│ └── rendered-html.test.mjs
├── AGENTS.md
├── CONTRIBUTING.md
└── LICENSE
Requires Node.js >=22.13.0.
npm install
npm run devBefore opening a pull request:
npm run checkVerify every pinned source against GitHub:
npm run sources:verify-remote- Source-backed founding index
- Search and stack filters
- Repeated-pattern view
- Indexable stack and pattern directories
- Stack-aware starter builder
- Responsive and reduced-motion support
- Expand to 100 reviewed sources
- Add automated upstream-change detection
- Publish an installable starter CLI
- Add measured before/after evaluations without replacing source evidence
Corrections and new sources are welcome. A valid source must be public, substantive, attributable, and pinned to a commit. Read CONTRIBUTING.md before submitting.
Source owners may request a correction, clarification, or removal at any time.
The gallery code and original project documentation are available under the MIT License.
The indexed AGENTS.md files remain governed by their source repositories.
This project links to source files by default and does not relicense them.