AGENT SKILL

outcomes

Use when work needs explicit definitions of done or an outcome grader should verify completion; defines measurable end states, gradeable criteria, required evidence, and pass or failure conditions.

License 许可
MIT
所属仓库
amoai-tech/mdeai
最近更新
2026-11-13
Stars
★ 2
Forks
1

01 前言 Front Matter

name
outcomes
description
Use when work needs explicit definitions of done or an outcome grader should verify completion; defines measurable end states, gradeable criteria, required evidence, and pass or failure conditions.

02 资源文件 1

03 SKILL.md 正文

OUTCOMES — Define & Grade Done

Goal Define the observable end state before execution, then verify it with explicit evidence.

1. State the outcome Describe the final result, not implementation steps. Good:

A renter submits one viewing request and exactly one lead/showing is persisted for the correct broker. Bad: Create an RPC and update the frontend.

2. Identify what is graded Specify the artifact or workflow:

  • code/change
  • file
  • deployed feature
  • database state
  • E2E journey
  • research/migration result

3. Write gradeable criteria Each criterion should be independently checkable. Prefer exact states, values, persisted results, authorization outcomes, and source-backed facts. Avoid “looks good,” “high quality,” or “production ready.”

4. Require evidence For important criteria define proof:

  • run exact test
  • inspect persisted row/state
  • verify deployed SHA
  • fetch cited source
  • prove unauthorized failure
  • prove retry creates no duplicate Do not pass on plausible-looking output alone.

5. Anticipate false success Ask how the agent could appear done while failing the real outcome. Examples:

  • UI says success but DB write failed
  • citation does not support claim
  • preview works but production is stale
  • retry duplicates transaction

6. Define what to ignore Exclude unrelated style issues, pre-existing defects, optional improvements, and unrelated warnings. Only grade the requested outcome.

7. Set iteration rules Typical:

  • simple: 2–3 iterations
  • moderate: 3–5
  • complex: up to 5–8 when justified Stop when all criteria pass, a real blocker exists, or max iterations is reached.

Output

Outcome One concise end-state statement.

For remaining steps, checks, and examples, see full prompt.

04 同仓库更多 Skills