Three short chats beat one mega-prompt: draft → critique → revise
Goal: Split draft / hard critique / revise into three handoff prompts—one job per turn.
Best for: Emails, posts, study notes, landing blurbs, stakeholder updates.
One mega-prompt that says “write and make it perfect” invites confident mush. Chain three moves: draft only → critique with a rubric → revise from the fix list.
Chaining separates roles: draft without judging, critique without rewriting, revise against a checklist. One blob prompt mixes them and hides failures.
Why beginners skip this
Treating one-shot generation as final feels light—without a critique chain, errors stay. Draft → critique → revise.
Step 1 — Draft only
Draft only: use real task material, mark gaps [TODO], and do not polish yet.
Goal: Draft v1 only — do not polish.
Format: [shape]
Boundaries: If info is missing, use [TODO] instead of inventing.
Input: [paste]
Step 2 — Critique with a rubric
Critique the v1 draft against the rubric below—scores and a fix list only; no rewrite yet.
Critique the draft against:
1) Clarity 2) Specific next step 3) No invented facts 4) Tone fit for [audience]
Return: score /10 each + fix list. Do not rewrite yet.
Example
Filled prompt
Live apology email: 3-day warehouse delay; offer CODE-WAIT only if already approved in notes.
Chain: outline → draft → critique (flag invented refunds) → revise.
Stop before send. I add one human detail myself.
What good output looks like
- Three separate prompts labeled Draft / Critique / Revise
- Critique prompt embeds your checklist (length, ask, banned filler)
- Revise prompt says “only fix what critique flagged”
- Each stage has a clear input and output artifact
- No single message tries to do all three jobs
Step 3 — Revise from the list
Revise only what the critique list named; leave everything else untouched.
Revise using only the fix list. Do not add new offers. Return final + what changed.
You remain the editor-in-chief: reject fixes that invent facts to raise a score.
Save the chain as a template card
Name it draft-critique-revise-v1. After a month, add one rubric line that kept failing. Version up.
Turn my three-step chain into a versioned prompt card with changelog line.
Example
Filled prompt
Live ask: customer note about bug #4412 ETA; facts: in QA, no date promised.
Run chain: outline → draft → critique → revise. Stop before send. Add human detail myself.
What good output looks like
- Clean split of the blended mega-prompt into three turns
- Handoffs name what to paste forward (draft text, critique bullets)
- No leftover “be harsh and also fix” in the same turn
- Checklist items are observable, not vibes
- You could run the pack next week without rewriting
Save this card
Card: Draft, critique, revise in separate turns—handoffs beat one mega-slurry prompt. Keep this beside the tool; read before you start.
Before you start
Keep the last failing output visible. You will repair the prompt for “Three short chats beat one mega-prompt:” against that failure—not a fresh toy ask. Guide: prompt-chaining-draft-critique-revise. 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 run a real draft through three links: draft → critique list → revise; no “final” until critique passes.
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).
Critique needs a checklist
“Be harsh” is not a rubric. Give three to five observable checks (length, banned words, missing ask). The revise turn should quote those checks—not invent a new direction.
One job per turn
Draft turns produce text. Critique turns produce a checklist-scored punch list. Revise turns apply only that list. If a single message tries to do all three, split it before you hit send—blurred jobs hide which stage failed.
Example
Filled prompt
Chain for customer apology: outline→draft→critique invented refund→revise.
Facts: 3-day delay, CODE-WAIT approved. Stop before send.
What good output looks like
- Goal/Format/Boundaries all present for “Improve quality by separating creating from judgin”
- One example beats stacked IMPORTANT words
- Version/date kept when template changes
- Pass/fail: zero clarifying turns on live task
Red flag
One mega prompt for draft+critique+revise; no handoffs—split turns.
Pitfalls
| Pitfall | Fix |
|---|---|
| Critique+rewrite in one ask | Split steps |
| Accepting all-10 vanity scores | Demand tough critique |
| Revising without a fix list | Step 2 mandatory |
Example
Filled prompt
Temptation: skip the habit for prompt-chaining-draft-critique-revise to save five minutes.
Job still: Improve quality by separating creating from
Smallest correct pass: Done visible · checkable artifact · Unknowns [?].
Boundaries: draft only; no silent send; no invented metrics.
Return: artifact + refuse-line said aloud.
What good output looks like
- Names the shortcut refused (specific to prompt-chaining-draft-critique-revise)
- Still delivers checkable output for “Three short chats beat one mega-prompt:”
- Unknowns labeled; draft-only
- Pass/fail: refuse-line dated (prompt-chaining-draft-critique-revise #4)
Practice now
Stop after one revise: second revise needs a new Goal (Asia/Shanghai).
Gate between: critique must produce a checklist before revise (Asia/Shanghai).
Chain drill: draft → critique checklist → revise once; no infinite loop (Asia/Shanghai).
Thin-spot drill: rewrite the Save card with real nouns from your work (repo, doc, inbox). Generic placeholders fail the drill.
Tonight run draft → critique → revise three rings on a real piece; save each ring’s artifact separately.
Buffer line: name the artifact path you used (Asia/Shanghai).
Ship-check: artifact path ___ · refuse-line dated Asia/Shanghai ___ · never-do ___
Close the loop
Write one line: next time I call one-shot final, I refuse.
One sentence takeaway
Draft → critique → revise; no “final” until critique passes.
Next
Rubrics · Voice pass · Outline first · Goal · Format · Boundaries
三段短对话胜过一条巨型提示:草稿 → 批评 → 修改
目标: 把「又要草稿、又要狠批、又要改好」拆成三条能交接的提示:每一段只干一件事。
适合: 单条巨型提示把三件事揉碎,输出忽好忽坏的人。
链式比吼一声「再优化」稳:先出草稿,再按清单批评,再只改被点名的地方。交接句写清楚,下一段才接得住。
为什么新手跳过
一次生成就当完稿像省事;缺批判链,错误会原样留下。草稿→批判→改写。
步骤 1——只起草
这一段只起草:用真实任务材料,缺信息标 [待办],先不要抛光。
目标:只要 v1 草稿,不要抛光。
格式:[形状]
边界:缺信息用 [待办],不要编。
输入:[粘贴]
步骤 2——按评分表批评
用刚写好的 v1 草稿,按评分表逐项打分;这一步先不要重写。
按以下标准批评草稿:
1)清晰 2)具体下一步 3)无编造事实 4)语气适合 [读者]
返回:每项 /10 分 + 修改清单。先不要重写。
实例
填好的提示词
为约 400 字的产品更新邮件建一条 3 段提示词链:(1) 草稿 (2) 按清单批评 (3) 修订。批评清单:不编指标、≤150 词、只有一个 CTA。
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。
像样的输出大概长这样
- 三条提示分开:(1) 草稿 (2) 按清单批评 (3) 修订
- 批评清单含:不编指标、≤150 词、一个 CTA
- 修订提示明确「只改批评点到的地方」
- 交接句清楚,下一条接得上上一条输出
- 不像一条巨型提示把三件事揉碎
步骤 3——按清单修改
只按批评清单修订:清单没点到的地方一律别动。
只按修改清单改。不要加新优惠。返回终稿 + 改了什么。
你仍是总编辑:拒绝为了提分而编造事实的「修复」。
把链条存成模板卡片
命名 draft-critique-revise-v1。一个月后,把总失败的一条评分加进去。升版本。
把我的三步链变成带变更说明的版本化提示词卡片。
实例
填好的提示词
我的单条提示词又要出草稿、又要狠批、又要改好。请拆成三条短提示,并写清彼此怎么交接。
批评用外部评分表;修订轮禁止碰清单没点到的地方。
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。
像样的输出大概长这样
- 拆分干净,草稿 / 批评 / 修订各管一段
- 每段写清要交给下一段的东西(全文 / 批评列表 / 定稿)
- 去掉「又要狠批又要一次完美」的矛盾指令
- 新手能按顺序复制三条去跑
- 总长度仍短过原来的单条巨兽
- 未知标出;不编造负责人、日期或指标
保存这张卡
记下:这篇对你有用的完成标准 + 绝不妥协的底线。标上日期。
开始前
把上一次失败输出留在眼前。没有失败样例的调试只是猜。
本周怎么用
本周对一篇真稿跑三链:草稿→批判清单→改写;批判不过关不宣布完稿。
把练习做完
用真实材料做完「现在就练一次」,再关标签页。
批评需要清单
「狠一点」不是量表。给出三到五条可观察检查(长度、禁用词、缺不缺请求)。修订轮应引用这些检查:不要另起炉灶。
一轮只干一件活
起草轮产出正文。批评轮产出按清单打分的修改点。修订轮只应用那份清单。如果一条消息想一次干完三件,发送前先拆开:职责糊了,就看不出哪一步失败。
红旗
一次生成就当完稿,缺批判链,错误会原样留下;草稿→批判→改写。
常见坑
| 坑 | 改法 |
|---|---|
| 一条提示又写又批又改 | 拆成三条并写清交接 |
| 批评没有清单 | 先定不编指标 / 篇幅 / CTA |
| 修订无视批评点 | 提示写「只改批评点到的地方」 |
现在就练一次
今晚就练:对一篇真稿跑三链:草稿→批判清单→改写;批判不过关不宣布完稿。
闭环
写下一行:下次一次生成就当完稿,我拒绝宣布。
一句话带走
草稿→批判→改写;批判不过关不宣布完稿。
下一步
← Back to home · Learn · Tools · Compare · Start