# Create Evidence-Based AGENTS.md Instructions

Missing or outdated repository instructions leave coding agents without reliable commands and editing boundaries. This skill drafts concise instructions from repository evidence and adds a CLAUDE.md bridge.

## Install

```bash
npx skillstore add agentkitco/write-agents-md
```

## Metadata

- Status: approved
- Slug: agentkitco-write-agents-md
- Version: 1.0
- Author version: 1.0
- Skillstore revision: r1
- Version status: invalid
- Tree hash: bf8ef31d00e47444f542efa5a9f2a4e05eedc6551bf54346163e113944baa7cc
- Author: agentkitco
- GitHub username: agentkitco
- License: MIT
- Repository: https://github.com/agentkitco/agentkit-dev-skills/tree/3de707c7875147b65ddff504d86b0d0dbf713fa7/skills/write-agents-md
- Ref: 743429f75b17204d47661a86b577a575ad67517f
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands
- Quality score: 75
- Public page: https://skillstore.pages.dev/skills/agentkitco-write-agents-md
- Manifest: https://skillstore.pages.dev/api/skills/agentkitco-write-agents-md/manifest

## Capabilities

- Inspect repository documentation, manifests, CI configuration, lint settings, and existing agent instructions.
- Draft AGENTS.md with commands, layout, conventions, boundaries, completion criteria, and open questions.
- Verify documented commands and paths while skipping destructive commands and reporting omissions.
- Create a CLAUDE.md import bridge or consolidate duplicated instructions in an existing file.
- Record missing evidence as owner TODOs instead of inventing project details.
- Keep instructions near 100 lines and move scoped rules into nested instruction files when needed.

## Use Cases

- Document a New Agent Workflow: Create initial repository instructions from existing documentation, scripts, and CI checks.
- Consolidate Team Instructions: Rewrite duplicated agent guidance into AGENTS.md and connect Claude Code through a shared import.
- Clarify Project Boundaries: Record generated files, approval requirements, test commands, and unanswered setup questions for contributors.

## Prompt Templates

### Create Initial Instructions

```
Create AGENTS.md for this repository using verified evidence. Add a CLAUDE.md bridge and list unresolved questions.
```

### Refresh Existing Guidance

```
Rewrite our AGENTS.md using current manifests and CI configuration. Verify safe commands, preserve project boundaries, and report skipped checks.
```

### Merge Tool Instructions

```
Review existing agent instruction files. Consolidate duplicated guidance into AGENTS.md, preserve unique rules, and update CLAUDE.md with an import.
```

### Organize Scoped Instructions

```
Create concise root instructions and scoped AGENTS.md files where needed. Verify paths and safe commands; flag unknowns and approval-required operations.
```

## Limitations

- Accuracy depends on accessible repository evidence; missing information remains an owner TODO.
- Command verification requires a suitable local environment and can execute repository-defined code.
- Destructive commands are skipped, so verification does not cover every documented operation.
- Tool compatibility depends on each agent's instruction-file support and configuration.

## Best Practices

- Use manifests, CI configuration, and existing documentation as evidence for every project-specific instruction.
- Review repository-defined commands before verification and keep destructive operations behind explicit approval.
- Keep shared instructions concise, preserve unique existing rules, and record unknowns as owner TODOs.

## Anti Patterns

- Inventing commands, paths, or conventions when repository evidence is missing.
- Copying secrets, personal paths, or duplicated guidance into instruction files.
- Running destructive commands during verification or committing documentation changes without a request.

## Security Audit

- Audited at: 2026-10-17T19:58:16.036\+00:00
- Summary: All 14 static findings are false positives: Markdown backticks format documentation, the environment-file reference prohibits committing secrets, and the personal-path reference prohibits disclosure. The workflow requests repository inspection, instruction-file edits, and command verification with explicit safety limits. No evidence found of prompt injection, credential collection, data exfiltration, or malicious execution in the two reviewed files.

## Stats

- Views: 0
- Downloads: 0
- Favorites: 0
- Popularity score: 0