write-goal

Help the user craft a well-specified goal objective for autonomous mode — turn a rough intention into a completion contract with a clear finish line, proof, boundaries, and stop rule. Use when the user asks for help writing, refining, or improving a goal, wants to set a structured objective for auto

Turn a rough user intention into a well-specified goal with a verifiable end state.

Anatomy of a Good Goal

Part	Description
Objective	One sentence: what will be true when done. Specific, actionable, verifiable.
Completion Criterion	How to check it's met. Concrete tests, commands, or observable outcomes.
Boundary	What is explicitly out of scope. Prevents scope creep.
Budget	Optional hard limit on turns, tokens, or time.

Workflow

1. Extract Intent

Identify from the user's request:

• What they want to achieve (the deliverable)
• Constraints — time, scope, tools, style

If already specific and verifiable, skip to step 3.

2. Identify Ambiguity

Gap	Action
No clear deliverable	Ask: "What should exist when this is done?"
No validation path	Ask: "How will you know it works?"
Too broad	Break into sub-goals or pick the most important slice
No scope boundary	Ask: "What is explicitly NOT part of this goal?"
Subjective quality	Replace with measurable criteria

If intent is clear from context, infer rather than ask.

3. Draft the Objective

Rules:

• One sentence, imperative mood
• Specific: name the deliverable, not the process
• Verifiable: completion criterion must be mechanically checkable
• Scoped: include boundary clause for broad requests
• Self-contained: a fresh agent with no context should understand it

Good:

text

"为 UserService 添加 JWT 认证中间件，实现登录/登出/token 刷新，所有现有测试保持通过。"
"Fix the memory leak in WebSocket handler — heap usage must stay stable over 10k messages."

Bad:

text

"改进代码质量"       — vague, no deliverable
"让页面好看一点"     — subjective, no validation

4. Define Completion Criterion

Prefer (in order):

1. Command: npm test passes, tsc --noEmit clean
2. File existence: specific file exists with expected content
3. Observable output: endpoint returns expected JSON
4. Manual check: only as last resort, must be specific

5. Commit the Goal

If goal management tools are available (e.g., CreateGoal, SetGoalBudget):

• Call the goal creation tool with the refined objective and completionCriterion.
• If the user specified a budget constraint, also set it:
  • "20 turns 内完成" → turns limit
  • "不要超过 500k tokens" → token limit
  • "30 分钟搞定" → time limit
• If an active goal already exists and the user wants to replace it, use the replace option.
• Do NOT set a budget unless the user explicitly asks.

If no goal tools are available:

• Present the structured goal as formatted text with clear sections: Objective, Completion Criterion, Boundary, Budget.
• Ask the user to confirm or adjust before proceeding with execution.

6. Present to User

Show the user:

• The objective (one line)
• The completion criterion (bullet list)
• Any budget set
• How to modify: "随时说 '修改目标' 或 '取消目标' 来调整"

Refinement

When improving an existing goal:

• If goal tools available: read the current goal state, identify weaknesses, propose a revised version, apply after approval.
• If no goal tools: ask the user for the current goal text, identify weaknesses using the anatomy table, show the diff, proceed after approval.

Multi-Goal Decomposition

For large requests:

1. Identify independent slices
2. Order by dependency
3. Focus on the first goal only — complete before next
4. Tell the user the full sequence

Anti-Patterns

Do NOT:

• Create goals for greetings, questions, or non-executable requests
• Describe the process instead of the outcome
• Use subjective completion criteria
• Set budgets the user didn't ask for
• Make the objective so narrow it's a single tool call

Language

Match the user's language. Technical terms and code identifiers stay in original form.