Agent conventions

Agent Conventions

Apply these conventions when creating or modifying AI agent definitions.

Dispatch

$ARGUMENTS	Action
Active (auto-invoked when working on agent files)	Apply all conventions below
Empty	Display convention summary
check	Run validation checks only

References

File	Purpose
references/readme-template.md	Template for agent README.md index entries

Scaffolding

Create new agent definitions with:

python skills/agent-conventions/scripts/scaffold_agent.py <name>

Conventions

Required Frontmatter

Every agent file must include these fields in YAML frontmatter:

• name -- kebab-case, must match filename without .md
• description -- non-empty, describes the agent's purpose

Optional Frontmatter

• tools -- comma-separated tool allowlist (default: all)
• disallowedTools -- comma-separated tool denylist
• model -- sonnet | opus | haiku | inherit (default: inherit)
• permissionMode -- default | acceptEdits | delegate | dontAsk | bypassPermissions | plan
• maxTurns -- integer cap on agentic turns
• skills -- list of skills preloaded into agent context
• mcpServers -- list of MCP servers available to this agent
• memory -- user | project | local
• hooks -- lifecycle hooks scoped to this agent

Memory field conventions:

• Choose the narrowest memory scope: local for temporary insights, project for shared patterns, user for cross-project preferences
• Keep memory entrypoint files under 200 lines — content beyond this is truncated at load time
• Organize memory by topic (e.g., debugging.md, api-patterns.md), not chronologically
• Use MEMORY.md as the index file; create separate topic files for detailed notes

Naming

• All agent names use kebab-case: ^[a-z0-9][a-z0-9-]*$
• Filename (without .md) must match the name frontmatter field exactly
• No consecutive hyphens, no leading/trailing hyphens

README Index Requirement

When defining a new agent at any level:

• ~/.{gemini|copilot|codex|claude}/agents/
• .claude/agents/ (project-level)
• Project-local agent directories

Update the corresponding README.md index in the same directory:

1. Add a row to the index table with agent name, description, and key fields
2. Add a description section with usage details
3. Keep the table sorted alphabetically by name

Body Content

• Write the system prompt in imperative voice ("Check the logs" not "Checks the logs")
• Keep agent definitions focused on a single responsibility
• Reference skills by name rather than inlining their content

Critical Rules

1. Always update the README.md index when adding or modifying an agent
2. Name every agent in kebab-case matching its filename exactly
3. Include both name and description in frontmatter -- they are required
4. Never duplicate agent functionality -- check existing agents first
5. Keep agent system prompts under 500 lines for maintainability
6. Run python scripts/validate_agent.py --check-index after any agent frontmatter change
7. Use imperative voice throughout the agent body text

Canonical terms (use these exactly):

• agent -- an AI agent definition file in agents/ directory
• frontmatter -- YAML metadata between --- delimiters
• system prompt -- the markdown body after frontmatter
• index table -- the markdown table in README.md listing all agents
• kebab-case -- lowercase words separated by hyphens

Validation Contract

Before declaring changes to this skill complete, run:

python skills/agent-conventions/scripts/check.py

Completion criteria:

1. scripts/check.py exits 0.
2. Agent frontmatter changes pass scripts/validate_agent.py --check-index.