.agents Guidelines
Guidelines specific to the .agents standard, extending the general guidelines.
Content
- Optimize for LLM context windows. Content in
.agentsis primarily consumed by AI agents. Write with token efficiency in mind — be direct, avoid redundancy across files, and structure content so it can be injected into context windows effectively. - Be factual and current. Profile and organization data should reflect the present state. Remove or update outdated information promptly rather than accumulating stale entries.
- Keep organization files user-scoped. Files in
profile/organizations/describe your relationship to an organization — your role, responsibilities, and current work. They are not canonical organization profiles; those live elsewhere (e.g., in the system-wide AgentFinger directory or the organization’s own documentation). - Scope techdocs to cross-project concerns. Techdocs cover tooling preferences, deployment strategies, language conventions, and other knowledge that spans multiple projects. Project-specific documentation belongs in the project repository (e.g., in a
CLAUDE.md,README.md, or.forge/).
Structure
- Keep
profile/flat. The profile directory has a fixed structure:user.mdat the top level and organization files inorganizations/. Do not create additional subdirectories. - Limit
techdocs/nesting. Use at most one level of subdirectories for categorization (e.g.,techdocs/languages/go.md). Deeper nesting makes discovery harder for tools scanning the directory. - Use descriptive file names. File names should clearly indicate their content:
containerization.md, notdoc1.md. The file name is often the only context a tool has before deciding whether to read the file.
Security
- Never store secrets. No API keys, tokens, passwords, or credentials.
.agentsdirectories can be served over the network via the AgentFinger protocol, so treat every file as potentially public. - Be deliberate about personal information.
profile/user.mdshould contain what’s useful for AI agents to personalize their assistance, not a full biography. Consider what you’d be comfortable with any tool on your machine — or on the network — reading.
Tool integration
Most AI coding assistants support instruction files (e.g., CLAUDE.md, .cursorrules, .windsurfrules). Use these to point tools at your .agents directory:
## Additional context
Read the following files before starting work:
- `~/.agents/profile/user.md` — user profile and preferences
- `~/.agents/techdocs/` — cross-project technical documentation
This ensures tools discover your .agents data without requiring built-in support for the convention.