One edit pass for voice consistency
Goal: Strip hot/cold AI voice from a draft without changing facts or inventing new promises.
Best for: After you already have a full draft (email, post, landing, update).
Voice drift looks like this: paragraph one is stiff corporate; paragraph two is sudden slang; paragraph three piles on fake empathy (“I totally understand how frustrating…”). One dedicated edit pass fixes more than five “make it better” chat turns.
An edit pass for voice catches tense shifts, suddenly formal words, and AI-favorite fillers. One pass with a checklist beats three vibe reads.
Voice is pattern, not perfume. If every sentence suddenly sounds like a brochure, the edit pass failed—even if grammar improved.
Why beginners skip this
Skipping a read-aloud after edits feels like one less step—voice jumps tire readers. Lock the voice, then edit.
Step 1 — Voice-only prompt
Paste a five-line voice card and tone-edit one section only—no new facts, offers, or apologies.
Edit for voice only.
Target voice: [plain / warm / formal]
Remove: hype adjectives, sudden slang, fake empathy lines, cheerleader openers.
Keep: facts, numbers I provided, structure.
Do not add offers, timelines, or apologies I did not write.
Return: revised text + 5 specific changes you made.
Step 2 — Customer email drill
Example
Filled prompt
Live artifact this week for: Remove AI-sounding tone swings across a draft—without changing facts or inventing new prom
Name real files/people. No invented facts; label Unknowns; stop before send/delete.
Return working prompt + one pass/fail check I can mark in a minute.
What good output looks like
- Edits target tense/formality/banned fillers only
- Change list names five concrete swaps (e.g. cut “delve”)
- Facts, numbers, and structure left untouched
- No new offers, apologies, or timelines inserted
- Paragraph still sounds like you—not a brochure
Step 3 — Stop when the change list repeats
If a second voice pass only renames the same five edits, stop—read aloud yourself instead of another model round.
Example
Filled prompt
Voice card: short sentences, ban "delve/robust/leverage", light warmth, no dunking.
Edit ONLY section 2 of this draft against the card. Return section 2 + 3 lines you removed.
Reject brochure openers.
What good output looks like
- Banned-filler list is personal (pulled from your pasted paragraph)
- Rewrite is shorter and blunter in your stated style
- Fake empathy lines (“I totally understand…”) removed
- No new claims, soft promises, or goodwill gifts added
- You could send after one human read-aloud
Save this card
Card: Voice card first, one section at a time—facts frozen, no new offers. Keep this beside the tool; read before you start.
Before you start
Name the reader and freeze facts on paper for “Remove AI-sounding tone swings across a.” If outline bullets do not exist yet, stop before prose. Guide: writing-edit-pass-voice-consistency. 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 lock three voice words on a real draft, fix inconsistent lines, read aloud before stop.
Voice-only pass
If facts moved during a “voice” edit, reject the turn. Tone is not a license to invent.
Voice is not a license to invent
If a “voice” edit moved a date, price, or promise, reject the turn. Tone changes words; it does not add goodwill gifts or new timelines. When the change list starts repeating, stop looping—read aloud instead.
Read aloud is the last gate
When the change list starts repeating the same five swaps, stop looping the model. Read the section out loud once. If a sentence still sounds like a brochure, cut it by hand—do not ask for another “One edit pass for voice consistency”
Example
Filled prompt
Voice card bans delve/robust/leverage.
Edit "Risks" section only; list removed words.
Use a live artifact from this week—not a toy paste. Label Unknowns; stop before send.
What good output looks like
- Matches outline/voice freeze for “One edit pass for voice consistency”
- No AI-slop openers; one concrete proof from your notes
- Numbers/names frozen — polish does not invent
- Pass/fail: phone-skim 20s still clear
Red flag
Editing without a voice card; new offers during polish—freeze facts, one section at a time.
Pitfalls
| Pitfall | Fix |
|---|---|
| “Make it better” with no voice target | Name plain / warm / formal |
| Letting the model add goodwill gifts | Ban new offers in Boundaries |
| Mixing voice edit with structure rewrite | Structure first (outline); voice second |
| No change list | Always ask for 5 concrete edits |
Example
Filled prompt
Temptation: skip the habit for writing-edit-pass-voice-consistency to save five minutes.
Job still: Remove AI-sounding tone swings across a
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 writing-edit-pass-voice-consistency)
- Still delivers checkable output for “One edit pass for voice consistency”
- Unknowns labeled; draft-only
- Pass/fail: refuse-line dated (writing-edit-pass-voice-consistency #4)
Practice now
Sample paragraph: lock voice from one golden paragraph (Asia/Shanghai).
Quote voice: quoted speech can differ; narration stays locked (Asia/Shanghai).
You/we lock: pick you or we; fix every mismatch (Asia/Shanghai).
Read-aloud cut: cut any sentence you stumble reading aloud (Asia/Shanghai).
Register lock: pick formal/casual once; flag every break in the margin (Asia/Shanghai).
Voice drill: mark three voice breaks in a draft; rewrite to one register (Asia/Shanghai).
Thin-spot drill: rewrite the Save card with real nouns from your work (repo, doc, inbox). Generic placeholders fail the drill.
Tonight lock three voice words on a real draft, fix inconsistent lines, read aloud before stop.
Ship-check: artifact path ___ · refuse-line dated Asia/Shanghai ___ · never-do ___
Close the loop
Write one line: next time I send edits with no read-aloud, I refuse.
One sentence takeaway
Lock three voice words first; fix inconsistent lines; read aloud before stop.
Next
Reject AI slop · Chaining · Stakeholder update · Outline before prose
一轮编辑:把语气统一
目标: 去掉草稿里忽冷忽热的 AI 腔,同时不改事实、不编造新承诺。
适合: 已经有完整草稿之后(邮件、帖子、落地页、进展更新)。
语气漂移通常长这样:第一段公文腔,第二段突然俚语,第三段堆假共情(「我完全理解您有多崩溃…」)。专门开一轮语气编辑,往往比五轮「再优化一下」更有效。
这一轮能抓住时态跳变、突然正式的词,以及 AI 爱用的填充词。带着清单改一轮,胜过凭感觉读三遍。
语气是模式,不是香水。如果每句突然都像说明书,哪怕语法更好,这一轮也算失败。
为什么新手跳过
改完不读出声像省一步;语气跳来跳去读者会累。先定声音再改。
步骤 1——只改语气的提示词
贴一张五条声线卡,只改一节语气;禁止新事实、优惠或道歉。
只编辑语气。
目标语气:[直白 / 温和 / 正式]
删除:吹嘘形容词、突然俚语、假共情、啦啦队开场。
保留:事实、我提供的数字、结构。
不要添加我没写的优惠、时间表或道歉。
返回:改后正文 + 你做的 5 处具体改动。
步骤 2——客户邮件演练
实例
填好的提示词
只做语气一致性编辑(不改事实)。清单:时态、正式度、禁用填充(delve、tapestry)、保持我的短句。草稿:【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。
像样的输出大概长这样
- 通篇人称与敬语一致,你 / 您不混用,读者称呼不跳来跳去
- 删掉连续三个「此外」「同时」把句子拐弯的地方
- 把被动堆砌改成主动短句,读起来更像人在说话
- 保留你口头禅里真正有辨识度的一两处,不洗成白板
- 改完后大声朗读一遍,不该在中途卡壳
步骤 3——改动列表开始重复就停
如果第二轮语气编辑只是把同一五处换个说法,就停;改成自己大声朗读,别再开一轮模型。
实例
填好的提示词
请从这段「很 AI」的段落生成个人禁用填充列表(≤8 条),并用我的直白话重写:【粘贴】。
禁止新事实、新优惠;一次只改语气。
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
- 对照表三列:原句 → 改句 → 改的原因(语气 / 节奏 / 事实)
- 这一遍只改声音与节奏,不偷偷加新卖点或新数据
- 至少三处示范,分别覆盖开头、中段、结尾
- 口吻像编辑批注,不像把全文重写一遍
- 下一稿还能拿同一张表自己再检一次
保存这张卡
记下:编辑轮或译→本地化里管用的一句改法。标上日期。
开始前
先备好读者是谁,以及一张冻结事实清单。如果不知道谁会读,先停:润色补不出读者。
本周怎么用
本周对一篇真稿定声音三词,通读改不一致句;读出声再停。
只改语气轮
如果「改语气」时事实动了,驳回这一轮。语气不是编造许可证。
语气不是编造许可证
如果「改语气」动到了日期、价格或承诺,整轮作废。语气改用词,不加赠品或新时间线。改动清单开始重复时就别再循环:改成朗读。
朗读是最后一道门
改动清单开始重复那五处替换时,就别再让模型循环。把该节朗读一遍。如果句子仍像宣传册,用手删:不要再要一轮「改语气」。
红旗
改完不读出声,语气跳来跳去读者会累;先定声音再改。
常见坑
| 坑 | 改法 |
|---|---|
| 语气轮顺手改事实 | 明确「只改语气,事实不动」 |
| 一次大抛光揉成淡语气 | 按清单逐节改,留棱角 |
| 禁用填充没入库 | 生成个人禁用列表并置顶 |
#今晚就练:对一篇真实稿先锁事实,再只开语气轮;发现改数字就停。轮;发现改数字就停。
闭环
写下一行:下次改完不读出声就发出,我拒绝。
现在就练一次
今晚就练:取一段自己的真实草稿,只跑一轮「语气一致」编辑;事实冻结不动。过/不过与永不做注明日期(Asia/Shanghai)。只收藏标题算失败。
一句话带走
声音三词先定;不一致句改完,读出声再停。
下一步
主动干掉 AI 套话 · 草稿链 · 干系人更新
← Back to home · Learn · Tools · Compare · Start