返回提示词目录第 1414 页

评估编程任务的智能体委派方式

根据可验证性、改动范围与审查成本评估编程任务,给出委派模式、人机分工及必要前提。

完整提示词

<context>
Coding agents do well when success can be checked automatically, the change is local, the codebase has patterns to copy and the request is unambiguous. They do badly on open design questions, changes that cross many modules with weak tests, work that needs knowledge nobody wrote down, production access, and security-sensitive code where a plausible mistake is costly. Delegation only pays when reviewing the result costs less than doing the work. A task that fails the check as a whole often passes once split: the human makes the design decision, the agent does the mechanical part.
</context>

<task>
Assess this task:
<task_text>
{{task}}
</task_text>
{{#context}}
<context_notes>
{{context}}
</context_notes>
{{/context}}

1. Rate each factor low, medium or high, with the evidence from the task or context (or "unknown"):
   - Verifiability: is there an automatic check that proves it works?
   - Locality: how many files and modules, and are interfaces changing?
   - Ambiguity: are there decisions left open?
   - Tacit knowledge: does it depend on unwritten history, customers or other teams?
   - Blast radius: data, money, security, public APIs or production effects if wrong.
   - Review cost: how long would a careful review of the likely diff take compared with doing it?
2. Decide: unattended (agent works alone and opens a pull request), supervised (agent works in a session with plan approval and checkpoints), split (name the human part and the agent part), or by hand.
3. Say how to run it in that mode: what the agent gets, the check that defines done, where a human approves.
4. List what stays human regardless (decisions, approvals, sensitive steps).
5. List prerequisites that would move the task to a more autonomous mode, such as a failing test that reproduces the bug or a decision recorded in the ticket.
</task>

<constraints>
- Unknown factors count against autonomy; do not assume tests exist.
- Do not recommend unattended mode for anything touching authentication, payments, data deletion or production infrastructure without a human approval step.
- Keep it short: this is a triage call, not a plan.
</constraints>

<output_format>
## Verdict
Unattended, supervised, split or by hand, with the deciding factor.
## Factors
Table: factor, rating, evidence.
## How to run it
Three to six bullets.
## Keep human
Bullets.
## Prerequisites
What would raise autonomy, or "None".
</output_format>

可替换内容

  • {{task}}
  • {{#context}}
  • {{context}}
  • {{/context}}

分类与标签

编程开发、编程智能体、任务委派、人机分工、代码审查、自动化评估

前往 AI 图片创作