trae-agent-writer（特工作家）

Trae Agent Writer

Create agent definitions for independent, isolated execution with business context.

What is an Agent?

• Spawned as subagents - Run with isolated context
• Single-purpose - One agent, one job
• Stateless - No memory between invocations
• Composable - Orchestrated by parent agents/skills

Phase 1: Understand Project (REQUIRED)

Before creating ANY agent, understand the context first.

1.1 Check Scope

If project is too large:

• ASK - What specific task needs an agent?
• SCOPE - Focus on one isolated task

1.2 Scan Existing Patterns

Quick scan to understand what exists (NOT deep reading):

1. Check if agents/ directory exists
2. List existing agents (names only)
3. Note invocation patterns used

Note: Deep reading of existing agents happens in Phase 2 when creating similar ones.

1.3 Understand Business

Agents need domain knowledge to make decisions:

Ask: "What criteria should this agent use?"

Phase 2: Create Agents (SEQUENTIAL)

Create agents ONE at a time.

2.1 Plan Agent Breakdown

First, identify what agents are needed:

Example: code-review skill
├── review-grader.md    (grade review quality)
├── code-comparator.md  (compare two versions)
└── issue-analyzer.md   (analyze patterns)

2.2 For EACH Agent

┌─────────────────────────────────────────────┐
│  For each agent:                            │
│                                             │
│  0. Initialize from template via script     │
│     - Run scripts/init_agent.py             │
│     - Edit generated scaffold               │
│                                             │
│  1. Define role clearly                     │
│     - Single purpose                        │
│     - What makes it need isolation?         │
│                                             │
│  2. Specify inputs/outputs                  │
│     - All parameters documented             │
│     - Structured output format              │
│                                             │
│  3. Write process steps                     │
│     - Numbered, clear steps                 │
│     - Include business rules                │
│                                             │
│  4. Move to next agent                      │
└─────────────────────────────────────────────┘

Initialization command:

python {skill_dir}/scripts/init_agent.py \
  --skill-dir {skill_dir} \
  --name {agent_name} \
  --role "One-line role" \
  --output-dir {project_root}/agents

2.3 Agent Format

# {Agent Name} Agent

{One-sentence role}

## Role

{What this agent does and why it needs isolation}

## What This Agent Should NOT Do

- ❌ **Do NOT {negative_action_1}** - {explanation}
- ❌ **Do NOT {negative_action_2}** - {explanation}
- ❌ **Do NOT {negative_action_3}** - {explanation}
- ❌ **Do NOT run commands or modify files** - Stay strictly read-only (unless explicitly a writer agent)
- ✅ **Only output**: {list_allowed_outputs}

## Inputs

- **param_name**: Description
- **output_path**: Where to save results

## Process

### Step 1: {Action}
1. Do this
2. Then this

### Step N: Write Results
Save to `{output_path}`.

## Output Format

{JSON structure}

## Guidelines

- **Be objective**: Avoid bias
- **Cite evidence**: Quote specific text

Phase 3: Quality & Lessons Learned

⚠️ Common Mistakes (CRITICAL)

These mistakes break agents. Always check:

Path Rule: Use relative paths like src/file.ts. For dynamic paths, use input parameters like {output_path}.

Quality Checklist

Before creating each agent:

• Paths - Use placeholders, no absolute paths
• Naming - Descriptive, action-oriented
• Language - Single language throughout
• Role - Clear single purpose
• Negative Constraints - "What This Agent Should NOT Do" section included
• Inputs - All parameters documented
• Output - Structured format defined
• Business - Includes domain context

Best Practices

Naming

Agent Locations

Good Agent Candidates

Don't make agents for: Simple inline tasks, tasks needing conversation history.

Example

User: "Create agent to grade code reviews"

Phase 1: Understand
- Purpose: Evaluate reviews objectively
- Needs isolation: Prevent bias
- Criteria: completeness, accuracy

Phase 2: Create

📄 agents/review-grader.md

# Review Grader Agent

Grade code reviews against quality expectations.

## Role

Assess reviews for completeness, accuracy, helpfulness.
Operates blindly to prevent bias.

## Inputs

- **review_path**: Path to review file
- **expectations**: List of expected findings
- **output_path**: Where to save grading.json

## Process

### Step 1: Read Review
1. Read review file
2. Extract all claims

### Step 2: Check Expectations
For each expectation:
1. Search for evidence
2. Mark PASS/FAIL
3. Cite specific text

### Step 3: Write Results
Save to `{output_path}/grading.json`

## Output Format

{
  "expectations": [
    {"text": "...", "passed": true, "evidence": "..."}
  ],
  "pass_rate": 0.80
}

## Guidelines

- **Be objective**: Don't favor verbose or brief
- **Cite evidence**: Quote specific text

Phase 3: Verify agent can be invoked

References

• Trae Agent Documentation
• Agent Patterns
• Grader Example