Recently Written · git

ai-config

One canonical AI agent config (rules, skills, MCP servers) rendered into Claude Code, Codex, Copilot CLI and Antigravity by acsync

git clone https://github.com/equwal/ai-config

Log | Files | Refs


skills/agent-creator/references/anthropic-agent-spec.md (3907 bytes)

1 # Anthropic Agent Spec — Official Reference
2 
3 Source: https://code.claude.com/docs/en/sub-agents (fetched 2026-04-05)
4 
5 ## Supported Frontmatter Fields
6 
7 Only `name` and `description` are required.
8 
9 | Field | Required | Description |
10 |:------|:---------|:------------|
11 | `name` | Yes | Unique identifier using lowercase letters and hyphens |
12 | `description` | Yes | When Claude should delegate to this subagent |
13 | `tools` | No | Tools the subagent can use (allowlist). Inherits all tools if omitted |
14 | `disallowedTools` | No | Tools to deny, removed from inherited or specified list |
15 | `model` | No | `sonnet`, `opus`, `haiku`, a full model ID, or `inherit`. Defaults to `inherit` |
16 | `permissionMode` | No | `default`, `acceptEdits`, `auto`, `dontAsk`, `bypassPermissions`, or `plan` |
17 | `maxTurns` | No | Maximum number of agentic turns before the subagent stops |
18 | `skills` | No | Skills to load into the subagent's context at startup (full content injected) |
19 | `mcpServers` | No | MCP servers available to this subagent (inline definition or string reference) |
20 | `hooks` | No | Lifecycle hooks scoped to this subagent |
21 | `memory` | No | Persistent memory scope: `user`, `project`, or `local` |
22 | `background` | No | Set to `true` to always run as background task. Default: `false` |
23 | `effort` | No | `low`, `medium`, `high`, `max` (Opus 4.6 only). Default: inherits from session |
24 | `isolation` | No | `worktree` — run in temporary git worktree |
25 | `color` | No | Display color: `red`, `blue`, `green`, `yellow`, `purple`, `orange`, `pink`, `cyan` |
26 | `initialPrompt` | No | Auto-submitted as first user turn when running as main agent via `--agent` |
27 
28 Total: 16 official fields.
29 
30 ## Key Facts
31 
32 - The **body** (markdown after frontmatter) becomes the **system prompt** that guides the subagent
33 - Subagents receive ONLY the system prompt + basic environment details, NOT the full Claude Code system prompt
34 - `tools` is an **allowlist** (like skills' `allowed-tools`)
35 - `disallowedTools` is a **denylist** — if both set, disallowed applied first, then tools resolved
36 - Naming parallel (marketplace note): agents use camelCase `disallowedTools`; skills use kebab-case `disallowed-tools` (schema 3.7.0+). The validator rejects either mismatch — never copy-paste between agent and skill frontmatter without renaming
37 - Subagents **cannot spawn other subagents** (no nesting)
38 - Subagents **don't inherit skills** from parent conversation — must list explicitly via `skills` field
39 
40 ## Plugin Agent Restrictions
41 
42 Plugin agents (`plugins/*/agents/*.md`) do NOT support:
43 
44 - `hooks` — ignored when loading from plugin
45 - `mcpServers` — ignored when loading from plugin
46 - `permissionMode` — ignored when loading from plugin
47 
48 These are standalone-only features. If needed, copy the agent to `.claude/agents/` or `~/.claude/agents/`.
49 
50 ## Tool Scoping
51 
52 Subagents can use `Agent(type)` syntax to restrict which subagents they can spawn (only for main-thread agents via `--agent`).
53 
54 ## Skills Preloading
55 
56 ```yaml
57 skills:
58   - api-conventions
59   - error-handling-patterns
60 ```
61 
62 Full content of each skill is injected at startup. This is the inverse of `context: fork` in skills.
63 
64 ## Model Resolution Order
65 
66 1. `CLAUDE_CODE_SUBAGENT_MODEL` env var
67 2. Per-invocation `model` parameter
68 3. Subagent definition's `model` frontmatter
69 4. Main conversation's model
70 
71 ## Example Agent File
72 
73 ```markdown
74 ---
75 name: code-reviewer
76 description: Reviews code for quality and best practices
77 tools: Read, Glob, Grep
78 model: sonnet
79 ---
80 
81 You are a code reviewer. When invoked, analyze the code and provide
82 specific, actionable feedback on quality, security, and best practices.
83 ```
84 
85 ## Scope Priority (highest to lowest)
86 
87 1. Managed settings (organization-wide)
88 2. `--agents` CLI flag (current session only)
89 3. `.claude/agents/` (project)
90 4. `~/.claude/agents/` (personal)
91 5. Plugin `agents/` directory (where plugin enabled)