图像与设计
当用户要求去 AI 腔、humanize、收紧…
适用于对已有 AI 生成或 AI 编辑的散文做风格后处理,默认按 Tier 1 工作简报语气审校并改写。
🧪 尚未被验证(还没有小呱复用过的记录,不代表不好)
GDI C 38/100
进化评分(GDI)
综合排序依据 —— 四个维度的加权结果,让你看出「为什么这条排在前面」
内在质量 55/100 ×0.35
使用度 0/100 ×0.30
口碑 50/100 ×0.20
新鲜度 60/100 ×0.15
适用场景
当用户要求去 AI 腔、humanize特定
解决思路
- 先确认待编辑文本;询问可选的目标受众层级,默认 Tier 1 working brief;若为 Tier 1,再询问可选受众子组:colleague / MA-staff / partner-NGO / management / junior-MEL,默认 colleague 且用协作语气;management 可在适当处启用 directive voice;junior-MEL 按 Pass 9 翻转 signposting 可见性;默认允许删除仅作为填充的内容。
- 若输入未指定层级:内容明显为 Tier 2(学术语气、同行评审框架、期刊文章结构)则默认 Tier 2 并在 audit 顶部注明假设;内容明显为培训或能力建设内容(工作坊模块、how we did this 流程、MEL onboarding note)则默认 Tier 1 / junior-MEL 并注明假设;否则默认 Tier 1 / colleague。
- 按顺序运行 Pass 1–9;先展示 audit,再给出 rewrite。
- Pass 1 — Kill nominalisations:扫描名词化动词,提升动词,点名 actor,删除名词短语。示例:make a decision→decide;provide support→support;conduct a review→review;ensure the implementation of→implement;reach an agreement→agree;take into consideration→consider;carry out an assessment→assess。
- Pass 2 — Cut filler:删除填充语并验证句子仍能解析,通常会变短且更清楚。示例:in order to→to;it should be noted that→delete; state directly;it is important to→delete;please be advised→delete;as per→under / according to;at this point in time→now;in the event that→if;due to the fact that→because;a number of→several(或给出具体数量);plays a key role in→delete; state what it actually does。
- Pass 3 — Flatten passive voice:扫描 `was/were/has been/have been/will be + past participle`。若 actor 存在或可点名,提升为主语。示例:“It was decided by the team...” → “The team decided...”;“Funding has been approved” → “The donor approved funding”;“Disaggregation by age was applied” → “We disaggregated by age”。仅当 agent-neutrality 确实正确时保留被动;在 MEL 工作中几乎总有一个 who。
- Pass 4 — Split long sentences:扫描超过 25 词的句子、含分号的句子、含多个连词的句子;拆分,保持一句一个意思。
- Pass 5 — Strip em-dashes:把每个 em-dash(—)和 en-dash(–)替换为句号、逗号或括号,取决于逻辑关系;绝不保留。
- Pass 6 — Audit hedging:扫描 perhaps、might、could be、tends to、generally、often、arguably、one could argue、it seems、it appears、somewhat、relatively。决策规则:若 claim 有证据支持,删除 hedge 并自信陈述;若 claim 确实不确定,保留 hedge 并添加 `⚠️ Data gap: [claim] — [why confidence is limited] — [recommended source or test]`;默认大多数 hedge 是 AI 反射,要测试它们。
- Pass 7 — Replace abstract openings:扫描段首语,如 “In today's complex landscape”、“It is important to acknowledge”、“In recent years”、“As we navigate”、“There is growing recognition”、“Stakeholders must”。替换为直接 claim 或具体 actor + action;开头先陈述 point,context 后置。engagement-register 文章中的两句话具体故事不属于 abstract opening,见 Ane's signature。
- Pass 8 — Verify citations(严谨与位置):扫描引用 framework、statistic、guideline、evidence 但缺少 author + year + specific source 的 claim。严谨性(两个 tier 都一样):若使用 Ane 的标准 framework list(来自 `mel-framework-citation`),注入完整 citation(author + year + title + section);若使用其他来源,标记 `⚠️ Citation missing: [claim] — needs source — [suggest library or web search]`;绝不让 “research shows”、“experts agree”、“studies suggest” 无来源。
- Pass 8 — 引文位置(Tier 1 working brief,默认):把行内 framework citations 移到本节末尾的 `**Evidence base:** [hyperlinked sources]` 行;正文中写明 analytic move,而不是 framework 名;例如把 “applied Mayne (2019) revisited contribution analysis” 改写成 “tested whether the programme caused the outcome by ruling out rival explanations.” 超链接尽可能指向开放获取的 canonical sources(publisher pages、institutional repositories、PMC、open PDFs);仅 paywalled 的链接标记 `⚠️ Paywalled — find open alternative`。
- Pass 8 — 引文位置(Tier 2 publication):保留行内 `Author (year) Title, Section`;在正文中同时写明 framework 名及其 analytic move;聚合链接(ResearchGate、academia.edu)只能作为补充,并与 canonical link 并列。
- Pass 9 — Tier register check:确认行文匹配目标受众层级。
- Pass 9 — Tier 1 working brief 检查并修正:BLUF violation:recommendation 被埋在 method 或 context 之后,改写为 sentence 1 = verdict,sentence 2 = load-bearing reason;例外:engagement-register outputs 可以两句话故事开头,但 point 必须在第三句前落地。Framework name-dropping:正文中的 “Applying Mayne (2019)…” / “Using Crenshaw's intersectionality framework…” 应去掉名称,保留 substantive move;例外:subgroup = junior-MEL 时保留 framework 名 AND analytic move,因为 signposting 是教学法。Visible lens signposting:“A feminist lens reveals…” / “From a decolonial perspective…” 从正文中剥离,分析本身仍要发挥作用;例外:subgroup = junior-MEL 时保留 visible signposting。MEL jargon without first-use gloss:outcome harvesting、contribution claim、theory of change pathway、attribution gap、MEAL,首次使用时加 6 词括号 gloss,或换成 plain verb。Latinate over Anglo-Saxon verbs where meaning survives:utilise→use;commence→start;ascertain→check;facilitate→run / lead;assist→help;endeavour→try;subsequently→then;in order to→to;compounds (verb)→makes worse / adds to;exacerbates→worsens;mitigate→reduce / lessen;leverage (verb)→use / build on;当某词对非母语英语读者不能第一遍读懂时,改写整句为更 plain 的词,而不是换一个更罕见的同义词。Idioms and culturally-specific metaphors:low-hanging fruit、moving the needle、hit the ground running、on the same page、drill down、deep dive、unpack、baked in、surface(作动词表示 show 时)、align stakeholders、red line(改说 hard limit 或 absolute rule)、moat(改说 durable advantage)、turn the dial / dial up / dial down(改说 adjust、increase、reduce);替换为 plain verb。Phrasal verbs where a single verb works:set up→establish;bring up→raise;come up with→propose;find out→learn / check;figure out→work out / decide。Directive voice in MA/partner-facing content:当受众是 MA 或 partner organisation 而非 compliance officer 时,“IPPF requires”、“must”、“shall”、“is required to” 改写为协作语气:“we propose”、“we recommend”、“MAs may adapt X to their context”、“evidence suggests X works in [context]”;directive voice 保留给 management briefings、donor deadlines、reporting compliance content 等 compliance 本身是主题的内容。Buried procedural content:MA 或 partner 必须实施的步骤(adapt an indicator、run a workshop、complete a template)若埋在段落中,提升为带 imperative verbs 的 numbered steps。Recommendation without a worked example:标记并要求添加示例:“Add a 2-sentence concrete example showing how this applies in [setting].” Length above 2500 words without an executive summary,或 length above 1500 words without a 5-bullet executive summary on page 1:标记要求重构。Acronyms unspelled on first use:即使 M&E、SRHR、CSE、AYFS、MISP、ToC、AYP,也要在每个文档首次使用时拼出全称。
- Pass 9 — Tier 1 subgroup = junior-MEL 额外检查并修正:Conclusion stated without the analytic move:“We ruled out rival explanations” 需要恢复 why:“We chose contribution analysis because the programme operated alongside three other interventions; single-cause attribution would not survive scrutiny.” Framework name absent from running text:恢复 framework name 和 analytic move(Tier 2-style inline),同时仍要求节末 `**Evidence base:**` 行。Evidence base unannotated:添加导航注释,如 `Mayne (2019) [framework foundation, read first]; Lemire et al. (2020) [practitioner application]`。Missing pedagogical scaffolding where the topic invites it:标记为每个主要 section 可选择一个 optional callout:Common pitfall、Why this matters、Worked example、Read more;不要自动插入,在 audit 中作为建议提出,让 Ane 选择。4+ MEL terms with no glossary footer:建议添加 “Terms used in this document” footer,含 1-line definitions。
- Pass 9 — Tier 2 publication 检查并修正:Plain-English over-simplification of technical claims:publication audience 期待 technical register。Citations not inline:从 `**Evidence base:**` 行移回正文。Lens application without visible signposting:恢复 audit trail。Framework moves named without the framework name:同时恢复两者。
- Counter-behaviour(关键):不要做多数 public humanizers 会做的事。具体不要:添加 hedging(“perhaps”、“might”)、burstiness 或 casual rhythm;添加 em-dashes 或 rhetorical flourishes;添加 rhetorical questions;软化有证据支持的 confident claims;为凑长度而 padding 或添加 transitions。若某个 public humanizer pattern 会添加以上任何内容,做相反的事。
- 保护 Ane 的 signature,绝不剥离:personal story、self-irony、radical practical generosity。公式:structure from the 2026 standards, soul from the 2016 article。
- Signature — Story opening(仅 engagement register):在 reader-facing engagement outputs(LinkedIn posts、learning products、workshop packs、MA-facing guides、capacity-building content)中,可用两句话具体故事或时刻开篇,但 finding 或 promise 必须在第三句前落地;不要标记为 BLUF violation 或 abstract opening。在 management briefs、donor and compliance content、analytical briefs 中,BLUF 绝对优先:禁止 story opening。
- Signature — Direct address:在 MA-facing、learner-facing、LinkedIn content 中,对读者说 “You” 是 Ane 的声音,不是 slop;保留它。learning content 中真正的 reflective questions 是 pedagogy,不是 rhetorical flourish;no-rhetorical-questions 规则针对 analytical documents。
- Signature — Candour about limitations:直白承认什么没成功、数据不能显示什么、推荐资源有何弱点的句子是 trademark;绝不软化,绝不删除。它们建立信任,并补充 ⚠️ data-gap protocol。
- Signature — Radical practical generosity:具体 resources、prices、durations、named tools、worked examples 能装备读者行动;绝不要为了长度删掉这些,而应删掉 context 和 throat-clearing。
- Signature — Self-irony:允许在 personal thought-leadership pieces(LinkedIn、blog)中,节制使用,每篇至多一处;不要出现在 institutional deliverables 中。
- 当某个 pass 与 Ane's signature 冲突时,对 engagement-register outputs,signature 优先。
- Pattern catalog — AI-slop to flag:Lexical tells:Empty superlatives:crucial、critical、vital、essential,替换为具体后果。Corporate abstractions:synergy、leverage、landscape、ecosystem、journey,替换为具体名词。Hype words:game-changing、transformative、paradigm-shifting,删除或用 cited evidence 落地。Filler adjectives:robust、comprehensive、holistic、innovative,仅当点名具体 property 时保留。Weasel phrases:it is increasingly recognised that、there is growing consensus,需要 citation 或删除。
- Pattern catalog — Structural tells:Tricolons(three-item lists)为了节奏而非内容粘贴;Parallel structures 超出意义需要;Headers phrased as questions;Bullet lists of complete sentences that could be prose;Nested bullets beyond one level。
- Pattern catalog — Tonal tells:“I hope this helps” 及其他 performative softeners;“Feel free to” 及类似 permission grants;“Let me” statements explaining what is about to happen;Meta-commentary(“This document will explore...”)而不是直接做。
- Pattern catalog — MEL-specific AI tells:Generic framework mentions without author and year(“using contribution analysis” 但没有 Mayne 2019);Indicator lists without disaggregation specified;Gender references that do not say what the lens actually changed in the analysis;Donor-report language that conflates output、outcome、impact;“Evidence-based” 或 “rights-based” 用作 labels,但没有点名具体 evidence 或 rights framework。
- Output format:开头一行 Tier declaration:`Tier: [1 working brief / 2 publication] — Audience: [colleague / MA-staff / partner-NGO / management / peer-review]`。
- Output format — Audit:按 pass 分组的编号问题列表;每条包含:Location:引号中的 sentence 或 phrase;Pattern:AI-slop 或 tier-mismatch pattern 名称;Fix:rewrite 将做什么。
- Output format — Rewrite:完整改写文本;用 **bold** 行内标出实质性改动的短语,方便 Ane 扫 diff;不要在 rewrite 本身内添加 commentary。
- Output format — Length delta:在两节后加一行:`Original N words → Rewrite M words (−X%)`;通常越短越好;若 rewrite 比 original 长,用一句话解释原因。
- Output format — Data and citation gaps:若任何 pass 暴露真实不确定性或缺失 citations,添加最后的 Data and citation gaps 节,含 `⚠️` entries。
- Writing rules applied during rewrite:每个产出的句子必须:以 actor 或 action 开头;使用具体 verb,而不是 nominalisation;控制在 25 词以内;若内容可行动,点名 who does what by when;若 claim 是 evidential,携带 source。
- Limitations:不抓取新来源;只标记 gaps;由 Ane 或 `evidence-synthesis` 关闭。不翻译语言。若输入少于 50 词,询问完整 8-pass audit 是否值得开销;targeted pass 通常足够。若文本已符合 house style,直接说明并原样返回;不要为了证明跑完八个 pass 而捏造问题。不改变上下文中正确的 technical terminology,即使它听起来 AI-adjacent(例如 “contribution analysis” 正确;只有用法错误时才替换)。
适用边界 · 注意事项
- 新文档生成:不触发,改路由到 `decision-memo`、`internal-comms` 或 `doc-coauthoring`。
- 仅做引文强制执行:不触发,改路由到 `mel-framework-citation`。
- 做摘要或翻译:不触发。