Rubrics and acceptance tests for prompts

Goal: Tell the model how you will score it first—aim at the checklist, not at “feels fine.”

Best for: Recurring tasks where “looks fine” is not enough: apologies, briefs, status updates, study summaries.

A rubric turns taste into tests. Keep it short (4–6 checks). Ask the model to score itself, then revise until checks pass—and you still do the final human gate.

A rubric turns “looks good” into pass/fail. Write 4–6 observable checks before you generate—so you are not grading your own vibe.

Freeze the rubric before you generate. If you rewrite the checklist to match a draft you already like, you are not evaluating—you are decorating.

Tape the rubric where you can see it while the model streams. Changing scores after you fall in love with a draft is cheating.

Why beginners skip this

Editing with no rubric feels like feedback—both sides will disagree on “good.” Write the rubric first.

Step 1 — Tiny rubric template

Copy a tiny rubric and freeze it before you score—do not edit the bar after you like a draft.

Acceptance tests (must pass):
- [ ] Has sections [A/B/C]
- [ ] No invented numbers / URLs
- [ ] ≤ [N] words
- [ ] One concrete next step
- [ ] Tone: [plain/warm/formal]
Score each Yes/No. If any No, revise once and re-score. Do not add new facts to pass.

Step 2 — Customer apology rubric

Deliverable: Friday update ≤150 words.
Rubric: traffic light · ≤3 owned asks · no new metrics · no fluff opener.
Score a draft; list fails.

Step 3 — Save rubrics with templates

Store the rubric under the same versioned prompt card. When a new failure mode appears twice, add a checkbox.

Deliverable: Friday stakeholder update ≤150 words.
Write rubric: 4 pass/fail checks (traffic light, ≤3 asks with owners, no new metrics, no fluff opener).
Run once on a draft; mark fail reasons.

Save this card

Card: Turn feeling-words into yes/no tests so a stranger can gate your draft. Keep this beside the tool; read before you start.

Before you start

Keep the last failing output visible. You will repair the prompt for “Rubrics and acceptance tests for prompts” against that failure—not a fresh toy ask. Guide: prompt-rubrics-and-acceptance-tests. Keep the real job visible the whole time; a blank box means stop and fetch the artifact first.

How to use this week

This week write 3 scorable criteria + 1 acceptance test for a real task; no stop until they pass.

Close the week by quoting aloud the refuse line you actually used once. If you cannot, rerun Practice now and date the pass/fail (Asia/Shanghai).

Freeze then generate

Changing the rubric to match a draft you like invalidates the test. Paper first.

Keep it real

Run this guide’s Practice now on work you will actually ship this week—not a fictional scenario. Dated notes beat perfect theory.

Do not bookmark-only

Finish Practice now, then tick this guide on your path. Titles unread are not skills.

Close the loop

Write one line: next time I edit with no rubric, I refuse stop.

Deliverable: image prompt for docs header.
Rubric 4 checks: subject, style, ≥3 must-nots, one revise instruction.
Score my weak prompt, then rewrite to pass.

Red flag

Feeling-words as Done; stranger cannot gate the draft—convert to yes/no tests.

Pitfalls

Temptation: skip the habit for prompt-rubrics-and-acceptance-tests to save five minutes.
Job still: Tell the model how you will score the
Smallest correct pass: Done visible · checkable artifact · Unknowns [?].
Boundaries: draft only; no silent send; no invented metrics.
Return: artifact + refuse-line said aloud.

Practice now

Golden file: one golden output checked into Save folder (Asia/Shanghai).

Must-pass tag: mark must-pass checks with ★ on the rubric (Asia/Shanghai).

Weight note: mark which rubric check is must-pass vs nice (Asia/Shanghai).

Fail fixture: one intentional failing sample must trip the rubric (Asia/Shanghai).

Rubric drill: write pass/fail rubric with 3 checks; run once on a draft (Asia/Shanghai).

Thin-spot drill: rewrite the Save card with real nouns from your work (repo, doc, inbox). Generic placeholders fail the drill.

Tonight write 3 scorable rubric lines + 1 acceptance veto for a real task; no done until it passes.

Today card (copy) — prompt-rubrics-and-acceptance-tests: 1. Artifact I will produce: ___ 2. Pass/fail check: ___ 3. Refuse-line if tempted to skip verify/send: ___ (date: Asia/Shanghai) 4. One never-do from this guide: ___

Extra pass: do the card on a real file today—reading alone does not count.

Ship-check: artifact path ___ · refuse-line dated Asia/Shanghai ___ · never-do ___

One sentence takeaway

3 scorable criteria + 1 acceptance test; no stop until pass.

Next

Chaining · GFB · Eval outputs

给提示词加评分表与验收测试

目标： 先告诉模型你将如何打分：让它对准清单，而不是对准「感觉还行」。

适合： 道歉信、简报、进展、学习摘要这类重复任务。

评分表把品味变成测试，保持短（4–6 项）。可以让模型先自评再改到全过，最终仍由你把关。生成前先把量规冻在纸上；为了迁就已经喜欢的草稿去改清单，那就不是评估，是装饰。

为什么新手跳过

没有评分标准就开改像有反馈；双方会对「好」各说各话。先写量规。

步骤 1——迷你评分表模板

先抄一张小量规，冻结后再打分；别看见喜欢的草稿再改标准。

验收测试（必须全过）：
- [ ] 含章节 [A/B/C]
- [ ] 无编造数字 / URL
- [ ] ≤ [N] 词
- [ ] 一个具体下一步
- [ ] 语气：[直白/温和/正式]
每项 Yes/No。如果有 No，改一轮再评分。不要为了过关加新事实。

步骤 2——客户道歉评分例

为客服邮件草稿做一张 6 条量规：语气、长度、不编政策、下一步清楚、不指责、需要则披露。给草稿每项打 1–5 分，并先列出失败项：【粘贴】。
边界：不编造；未知标 [?]；发送前停下。
完成检查：一分钟内能判过/不过。

步骤 3——和模板一起存

把评分表放在同一张版本化提示词卡片下面。同一种失败模式出现两次，就加一条勾选项。

把「专业又有帮助」收成陌生人也能用的 5 条内部周报验收测试（是/否）。
场景：周五状态邮件。禁止感觉词；每条要能打勾。冻结后再打分。
边界：不编造；未知标 [?]；发送前停下。
完成检查：一分钟内能判过/不过。

保存这张卡

记下：今天改管用的那条提示词（含边界）+ 版本或日期。

开始前

把上一份失败草稿留在眼前。没有失败样例，打分只是猜。

本周怎么用

本周给一个真任务写 3 条可打分标准 + 1 条验收测试；不过关不收工。

先冻结再生成

为了迁就已经喜欢的草稿去改量规，测试就作废了。先写在纸上。

保持真实

用本周真要交付的工作做「现在就练一次」，别编假场景。带日期的实操笔记，比完美理论有用。

不要只收藏

先做完「现在就练一次」，再在学习路径上打勾。只收藏不练，标题变不成技能。

闭环

写下一行：下次没量规就开改，我拒绝收工。

红旗

没有评分标准就开改，双方会对「好」各说各话；先写量规。

常见坑

现在就练一次

今晚就练：给一个真任务写 3 条可打分标准 + 1 条验收测试；不过关不收工。

一句话带走

3 条可打分标准 + 1 条验收测试；不过关不收工。

下一步

链式修改 · 目标·格式·边界 · 自评输出