Neo Sub-Agent

Use this skill to design and generate focused sub-agent definitions for developer CLIs. Keep the user in control of scope, permissions, and target clients.

Core Rule

Do not invent client-specific schema. If the requested CLI, field, or file location is not documented in references/client-adapters.md or discoverable in the project, state the missing fact and generate only a neutral blueprint.

Workflow

1. Perceive

1. Inspect the project before asking questions:

• Existing agent definitions: .claude/agents/, .codex/agents/, .github/agents/, .agents/skills/.
• Project guidance: AGENTS.md, CLAUDE.md, GEMINI.md, .github/copilot-instructions.md.
• Related skills or conventions in skills/, .agents/skills/, .github/skills/, .claude/skills/, .codex/skills/.

1. Identify the target clients from the user request or project evidence: claude, codex, copilot, agy, or multiple.
2. If multiple clients are plausible and the request does not specify one, ask one concise question before writing files.

2. Decide Whether a Sub-Agent Fits

Read references/sub-agent-design.md when the request involves architecture, multiple agents, or tradeoffs.

Use a sub-agent when at least one is true:

• The work produces verbose logs, search results, or file reads that should not pollute the main context.
• The work is a repeated specialist role such as code review, planning, test execution, research, migration analysis, or docs verification.
• The work can run independently or in a clearly bounded sequence.
• The role needs stricter tool access, sandboxing, model choice, or output contract than the main agent.

Prefer a normal skill or main-conversation instruction when:

• The task is a quick, targeted change.
• The agent needs frequent back-and-forth with the user.
• Multiple workers would edit the same files concurrently.
• The workflow depends on undocumented client behavior.

3. Define the Agent Spec

Collect or infer these fields, using safe defaults where reasonable:

• name: lowercase kebab-case, under 64 characters.
• description: trigger-focused; state exactly when to use the agent.
• instructions: role, workflow, constraints, and output contract.
• clients: target clients.
• scope: project by default; use user only when the user asks for reusable personal agents.
• tools: use the smallest useful set; prefer read-only for reviewers, researchers, planners, and auditors.
• model and reasoning effort: omit unless the user asks or the client-specific reference gives a clear default.
• handoff: what the parent agent should pass in and what the sub-agent must return.
• validation: commands, checks, or review criteria the sub-agent should run or report.

4. Generate Files

Read references/client-adapters.md before generating client-specific files.

For repeatable output, create a JSON spec and run:

uv run skills/neo-sub-agent/scripts/render-sub-agent.py --spec spec.json

If uv is not available and the script has no external dependencies, use:

python skills/neo-sub-agent/scripts/render-sub-agent.py --spec spec.json

Review the dry-run JSON. If the target paths and content are correct, write files:

uv run skills/neo-sub-agent/scripts/render-sub-agent.py --spec spec.json --write

Use --force only when replacing an existing agent file is explicitly intended.

5. Review

Before finalizing, check:

• The agent has one narrow job and does not duplicate an existing agent.
• The description is specific enough for automatic delegation.
• Tool permissions match the role and avoid broad write/shell access when unnecessary.
• The output contract is short, concrete, and easy for the parent agent to synthesize.
• Antigravity output is labeled as a skill/delegation blueprint, not a custom sub-agent manifest.

Output Format

When reporting back to the user, use Traditional Chinese (Taiwan):

## 已新增
- [client] `path/to/agent-file`

## 設計重點
- 角色：
- 觸發條件：
- 權限：
- 回傳格式：

## 驗證
- `command`

Constraints

• Do not create broad "do everything" agents.
• Do not give background agents write or shell access unless the task requires it.
• Do not create multiple agents that can edit the same file set in parallel.
• Do not treat model-generated review as a replacement for deterministic tests.