Skip to content

A local-first multi-agent app development platform, built on PenguinHarness

docs(skills): humanizer round-six tells from a user editing session (#275)

Yaowei Zhengf293471c3fd8
diff --git a/packages/skills/skills/humanizer/SKILL.md b/packages/skills/skills/humanizer/SKILL.md
index e909a787..17cfa523 100644
--- a/packages/skills/skills/humanizer/SKILL.md
+++ b/packages/skills/skills/humanizer/SKILL.md
@@ -1,16 +1,16 @@
---
name: humanizer
-description: Rewrite or edit prose in any language so it reads like edited human writing in the register of books, newspapers and encyclopedias rather than default AI output. A small drafting core — vary every pattern, build density from anchored facts in whole grammar, cap the quotables, put a real writer with real material behind the text, let structure serve content, write each language from inside its idiom and typography, verify what you assert, and aim for the natural distribution of edited prose rather than a perfect scorecard — backed by a three-layer tell catalog, per-language cue files for six languages and a five-round measured case study shipped as reference files for the diagnostic census.
+description: Rewrite or edit prose in any language so it reads like edited human writing in the register of books, newspapers and encyclopedias rather than default AI output. A small drafting core — vary every pattern, build density from anchored facts in whole grammar, cap the quotables, put a real writer with real material behind the text, let structure serve content, write each language from inside its idiom and typography, verify what you assert, and aim for the natural distribution of edited prose rather than a perfect scorecard — backed by a three-layer tell catalog, per-language cue files for six languages and a six-round measured case study shipped as reference files for the diagnostic census.
short_description: Strip AI-writing tells from prose in any language.
short_description_zh: 去除文本的 AI 味,把行文改成书籍、报纸、百科式的风格,任何语言通用。
preinstall: false
-version: 7
-updated: 2026-08-11T00:00:00Z
+version: 8
+updated: 2026-08-12T00:00:00Z
---
# Humanizer
-Make prose read like edited human writing: the register of books, quality newspapers and encyclopedia entries. Everything here was derived empirically, across five rounds of drafting, blind editorial review and revision, documented with counts in [`reference/case-study.md`](reference/case-study.md). The working surface is deliberately small: a writer boxed in by a long checklist produces compliance, not prose — that failure mode is the case study's best-documented finding. Draft with the principles below; diagnose with the catalog afterwards.
+Make prose read like edited human writing: the register of books, quality newspapers and encyclopedia entries. Everything here was derived empirically, across six rounds of drafting, blind editorial review and revision, documented with counts in [`reference/case-study.md`](reference/case-study.md). The working surface is deliberately small: a writer boxed in by a long checklist produces compliance, not prose — that failure mode is the case study's best-documented finding. Draft with the principles below; diagnose with the catalog afterwards.
## Before you start
@@ -26,7 +26,7 @@ If the invocation carries no text and no assignment, ask for one: the draft to e
6. **Write from inside the language.** Native idiom — a sentence that back-translates cleanly into another language was composed there, so recompose it — native punctuation at native frequency (a dash where the register expects one beats zero), and the venue's typography held consistently. Per-language budgets and surface forms: [`reference/language-cues.md`](reference/language-cues.md), which indexes one `reference/<lang>-cues.md` file per language.
7. **True, verified, calibrated.** Never invent specifics. Re-derive every mechanism example from the stated mechanism, split fused attributions, and verify or delete every superlative. One fluent falsehood outweighs any amount of style.
-Over all seven: **aim for the natural distribution of edited prose, not a perfect scorecard.** Every rule above, overdriven, mints a new tell — the case study documents three such artifacts (banned openers became cold-open verdicts; scrubbed punctuation became semicolon inflation and rationed warmth; density became telegraph compression). Keep some looseness, an aside, an unresolved edge.
+Over all seven: **aim for the natural distribution of edited prose, not a perfect scorecard.** Every rule above, overdriven, mints a new tell — the case study documents four such artifacts (banned openers became cold-open verdicts; scrubbed punctuation became semicolon inflation and rationed warmth; density became telegraph compression; scrubbed abstraction became contrived colloquialism). Keep some looseness, an aside, an unresolved edge.
## Method
diff --git a/packages/skills/skills/humanizer/reference/case-study.md b/packages/skills/skills/humanizer/reference/case-study.md
index f9f80557..febd4a09 100644
--- a/packages/skills/skills/humanizer/reference/case-study.md
+++ b/packages/skills/skills/humanizer/reference/case-study.md
@@ -1,6 +1,6 @@
# Case study: measuring the AI out of prose
-How the tells in SKILL.md were derived, in five rounds. Round one is a controlled exercise: one assignment answered twice, first as a deliberately typical AI draft, the register a default assistant produces unprompted, then as a rewrite into print register; both versions were kept, the tells were counted with pattern matching rather than judgment, and the diff produced the sentence-layer tells (1–10). Round two is a set of field tests, added after human readers flagged text that already followed round one; it produced the discourse-layer tells (11–17). Round three scaled the method: ten fresh pieces across five languages and ten genres, written honestly to the first two layers, were reviewed blind by independent editors; what they still caught produced the scrub-layer tells (18–24) and the per-language lists in `language-cues.md`. Round four is a user-run field test on fresh material that caught the density rule overshooting (tell 25), and round five a further user pass on regenerated material that caught the repair itself compressing (clipped conditionals, particle-chain monotony). Rounds one and two are in Chinese; round three covers Chinese, English, Japanese, French, German and Spanish.
+How the tells in SKILL.md were derived, in six rounds. Round one is a controlled exercise: one assignment answered twice, first as a deliberately typical AI draft, the register a default assistant produces unprompted, then as a rewrite into print register; both versions were kept, the tells were counted with pattern matching rather than judgment, and the diff produced the sentence-layer tells (1–10). Round two is a set of field tests, added after human readers flagged text that already followed round one; it produced the discourse-layer tells (11–17). Round three scaled the method: ten fresh pieces across five languages and ten genres, written honestly to the first two layers, were reviewed blind by independent editors; what they still caught produced the scrub-layer tells (18–24) and the per-language lists in `language-cues.md`. Round four is a user-run field test on fresh material that caught the density rule overshooting (tell 25), and round five a further user pass on regenerated material that caught the repair itself compressing (clipped conditionals, particle-chain monotony). Round six is a user editing session on one release-note passage whose four successive rejections (让-chains, announcement scaffolding, the before/after frame, contrived colloquialism) converged on plain statement. Rounds one and two are in Chinese; round three covers Chinese, English, Japanese, French, German and Spanish.
The assignment, as a user would phrase it: 请帮我写一篇图文稿,大约 1000 字,讲清楚"skill 是什么"。(Write an illustrated piece of about 1,000 characters explaining what a skill is.)
@@ -212,3 +212,19 @@ An explainer regenerated under the round-four rules drew two further flags from
34. **Compression re-emerges at ever smaller scales.** Year-stamp runs (round four) gave way to four-character clause pivots and single-particle logic chains; the census must follow the pressure downward — count the workhorse connective and the clipped conditionals, not just clause chains.
35. **A spot that survives repair is overloaded.** The same sentence failed in three different wordings across three rounds; no wording fixes a sentence asked to carry cause, scale and consequence in one breath. Local repair loops end by restructuring the thought, and the method's gate now says so.
+
+## Round six: user field test, the announcement frame
+
+A user brought one zh release-note passage and rejected four successive rewrites, each fix drawing the next flag. The original:
+
+> 前几个版本,我们让 Agent 能够读写文件、调用技能、启动子代理和执行定时任务。这一次,我们让 Agent 拥有了长期记忆,让它在新的会话中也能延续之前积累的信息。
+
+The first rewrite broke the 「让」字链 — three causative clauses on one engine — but kept the paired opener, and was rejected: 「前几个版本」「这一版」这种排比也很 AI 味。 The second dropped the parallelism for one-way progression, existing abilities first and 现在它又多了长期记忆 after, and was rejected too: 这种过去和现在的对比就很 AI 味,只写现在就行。 The third stated only the present but kept 照样用得上 — the first rewrite's patch for the abstract 延续之前积累的信息 — now flagged: 这种感觉很刻意。 Accepted, with the flatter 之前积累的信息也还在 offered as an equal:
+
+> Agent 有了长期记忆,新会话里也能接着用之前积累的信息。
+
+### Findings, round six
+
+36. **The announcement frame is a tell at every strength.** Full 排比 scaffold, then bare before/after progression with no parallel wording: each weakening of the frame was still flagged, and the accepted fix stated the present alone — the change is the news, the frame adds nothing. The temporal frame joined tell 1's contrast family with the same budget and fix; the zh cues gained the 「让」字链 and the 今昔架子.
+37. **Repair overshoots into performative casualness.** 照样用得上, installed as the fix for an abstract phrase, survived two further rounds before being flagged as deliberate: the fourth documented overcorrection artifact, after the cold open, the scrub fingerprint and telegraph compression. Tell 19 extended — when an abstract phrase needs replacing, the plainest statement beats the folksiest.
+38. **One session can run the whole arc.** 让-chain, 排比, contrast frame, contrived colloquialism: each round's fix minted the next round's tell, and four exchanges converged on the plain present-state sentence. The gate's natural-distribution reading is not only a cross-piece average; it is where a single sentence's repair loop ends.
diff --git a/packages/skills/skills/humanizer/reference/tells.md b/packages/skills/skills/humanizer/reference/tells.md
index 229f279c..36194820 100644
--- a/packages/skills/skills/humanizer/reference/tells.md
+++ b/packages/skills/skills/humanizer/reference/tells.md
@@ -1,12 +1,12 @@
# The tell catalog
-The diagnostic instrument behind the humanizer skill: twenty-five tells in three layers, used in the census and revision steps of the method — not as drafting constraints. Drafting against this list produces compliance-shaped prose; the scrub layer below exists because of exactly that. The counts and quoted patterns come from the four measured rounds in [`case-study.md`](case-study.md); per-language surface forms are in [`language-cues.md`](language-cues.md).
+The diagnostic instrument behind the humanizer skill: twenty-five tells in three layers, used in the census and revision steps of the method — not as drafting constraints. Drafting against this list produces compliance-shaped prose; the scrub layer below exists because of exactly that. The counts and quoted patterns come from the six measured rounds in [`case-study.md`](case-study.md); per-language surface forms are in [`language-cues.md`](language-cues.md).
## Sentence layer
Countable patterns. For each: what to look for, then the fix.
-1. **Template contrasts.** The correction-shaped sentence recurs: "not X, but Y", "It's not just X, it's Y" (en); 不是……而是……、这不仅是……更是…… (zh). The negation pivot belongs to the same family: "he was not trying to advance maritime technology; he wanted to stop paying by the hour" stages a strawman nobody holds just to swerve off it. All correction-shaped frames share one budget: at most one per piece, and it must correct something a reader might actually believe. Fix: state Y directly and let X go unmentioned, or concede X in a plain subordinate clause.
+1. **Template contrasts.** The correction-shaped sentence recurs: "not X, but Y", "It's not just X, it's Y" (en); 不是……而是……、这不仅是……更是…… (zh). The negation pivot belongs to the same family: "he was not trying to advance maritime technology; he wanted to stop paying by the hour" stages a strawman nobody holds just to swerve off it. All correction-shaped frames share one budget: at most one per piece, and it must correct something a reader might actually believe. Fix: state Y directly and let X go unmentioned, or concede X in a plain subordinate clause. The before/after frame is the family's temporal member: "Previously… Now…", 前几个版本……这一次/这一版…… stages the old state just to land the new one, and it survives its own repair — parallelism stripped, contrast softened to bare progression (本来就会 X,现在又多了 Y), it still reads as launch copy. Same fix: state the present and leave the past unmentioned (只写现在).
2. **Rhetorical triads.** Three parallel adjectives, three parallel clauses, three bullet virtues, whole sections in threes ("faster, safer, smarter"; 高效、可靠、优雅), and the vivid concrete triad with an expanding final item ("the sick week, the flaky test suite, the dependency upgrade that breaks the build"). Factuality is no defense at frequency: a field test found five three-part structures in 430 words, each defensible alone, unmistakable together — human prose uses the triad, it does not sustain one per paragraph. Census counts every three-part structure; budget roughly one per piece. Fix: keep the true items and break the symmetry: two items, four, or one item expanded with a fact.
3. **Dash overuse.** Em-dash asides every few sentences (en); 破折号 doing the work of commas and colons (zh). Print prose spends dashes rarely. Fix: a comma, a colon, parentheses, or a full stop and a new sentence. More than about one dash per several hundred words is a tell; the natural floor is genre- and language-relative (see the scrub layer).
4. **Decorative metaphors.** Figures that add mood, not meaning: "a digital symphony", "unlock the magic"; 像一位不知疲倦的助手、知识的魔力. Test a figure by asking what it lets the reader compute; if nothing, cut it. An analogy that compresses a mechanism may stay, at most one per piece; the round-one case study ended up keeping none.
@@ -34,7 +34,7 @@ A piece can score zero on every counter above and still read as AI. In field tes
The humanizing pass leaves its own fingerprint. In blind field tests, ten pieces written to the two layers above were scored 20–85% AI-likely by independent editors, and the strongest remaining signals were not classic AI patterns but the marks of their removal: "prose written under an anti-AI blacklist", as one review put it. A perfect scorecard is itself the giveaway; the target is the natural distribution of edited prose, not zero.
18. **Visible avoidance.** Zero dashes and zero parentheses in a register that breathes through both, semicolons and comma splices doing the banned marks' work ("did not improve the old system, they replaced it"), colons exterminated from stat-bearing prose, every ordinary contrast routed around so consistently the detour shows. Punctuation budgets are language-relative: Spanish expects an occasional inciso entre rayas, French its colon reveals, German its Gedankenstrich. Fix: restore the genre's natural mark distribution; one dash where the register expects one beats none.
-19. **Rationed humanity.** Colloquialism dosed one-per-paragraph like seasoning ("warmth applied with a dropper"), fake liveliness installed where clichés were removed (machine-gun 体言止め, fragments For Effect), first person amputated from opinion writing, anecdotes with no owner ("The feature was scoped at two weeks" — whose feature, where?). Fix: voice is a stake, not a garnish. Own the anecdote, risk one judgment of your own, let warmth cluster where the material warrants it and vanish elsewhere.
+19. **Rationed humanity.** Colloquialism dosed one-per-paragraph like seasoning ("warmth applied with a dropper"), fake liveliness installed where clichés were removed (machine-gun 体言止め, fragments For Effect), the abstract phrase patched with a folksy one (照样用得上 swapped in for 延续之前积累的信息, when the plain 之前积累的信息也还在 was wanted), first person amputated from opinion writing, anecdotes with no owner ("The feature was scoped at two weeks" — whose feature, where?). Fix: voice is a stake, not a garnish. Own the anecdote, risk one judgment of your own, let warmth cluster where the material warrants it and vanish elsewhere; when an abstract phrase needs replacing, the plainest statement beats the folksiest.
20. **Capsule airlessness.** Every paragraph a sealed module: one move, a hinge opener, a cadenced button, no digression, no stray fact, no dud — and the same architecture repeated across pieces (four paragraphs, scene lede, quiet kicker), the portfolio fingerprint of one process behind different masks. Fix: let one digression leak, let one paragraph end flat, vary paragraph count and opening attack between pieces.
21. **Genre without its material.** Reportage with no scene, no named person, no quotation; "journalism" whose respondents exist only as aggregates — what the never-invent rule produces when the genre needs material nobody gathered. Fix: get the material (research, interviews, the user's own facts) or relabel the piece as essay or explainer; never fake the texture.
22. **Canonical everything.** Every fact from the single standard account, exactly the two maximum-probability citations and no third, the default arc and the default title shape, the adjacent famous fact swapped in where the precise one should be. A researching human drags in one eccentric but true detail; retrieval regresses to the canon. Fix: add one off-canon specific you can verify, or narrow the claim to what the canon actually supports.
@@ -46,6 +46,6 @@ The humanizing pass leaves its own fingerprint. In blind field tests, ten pieces
Run these mechanically over the draft; impression is not a count.
-- Sentence layer: dashes; contrast templates, negation pivots included; every three-part structure; paragraphs of two sentences or fewer; hype words; scare quotes; reader addresses; subheads.
+- Sentence layer: dashes; contrast templates, negation pivots and before/after frames included; every three-part structure; paragraphs of two sentences or fewer; hype words; scare quotes; reader addresses; subheads.
- Discourse layer: restatements of the central thesis; turned phrases of every shape; sentences with a concept as subject; absolutes; adjacent sentences or clauses sharing one skeleton; clauses opening on a date; the workhorse connective's count per paragraph; clipped conditional pivots; semicolons against the language's budget; subjectless clauses whose referent lives elsewhere; and, outside English, calques that back-translate cleanly.
- Scrub layer: punctuation distribution against the genre's norm (too few is a signal too); spacing of colloquial warmth; citation profile (all canonical?); unverified superlatives; typography consistency; and, across several pieces from one session, shared architecture.
diff --git a/packages/skills/skills/humanizer/reference/zh-cues.md b/packages/skills/skills/humanizer/reference/zh-cues.md
index 38d9903c..cec2bb6f 100644
--- a/packages/skills/skills/humanizer/reference/zh-cues.md
+++ b/packages/skills/skills/humanizer/reference/zh-cues.md
@@ -3,6 +3,8 @@
Default-AI markers:
- 「不是……而是……」及其亲属(「不需要 X,只需要 Y」「这不仅是……更是……」)翻案句。
+- 今昔对比架子:发布腔的「前几个版本……这一次/这一版……」对仗开场;撤掉排比、改成「本来就会 X,现在又多了 Y」的递进,对比框架仍在、仍是 AI 味。只写现在,变化本身就是信息。
+- 「让」字链:连串使动句当引擎(让 Agent 能够……让 Agent 拥有了……让它在……);改成直陈(Agent 能……,它多了……)。
- 首尾升华:宏大开场(在……的今天)与拔高结尾(未来已来;……的故事仍在继续)。
- 三连排比与四字格堆叠(一种态度、一种选择、一种生活方式;源远流长、波澜壮阔)。
- 显性路标(首先/其次/最后、总而言之、值得注意的是)与设问自答开头。
@@ -19,6 +21,7 @@ Scrubbed-text markers (over-correction):
- 冒号、破折号、引语、史料味全部清零,逗号串承担一切(七个分句一逗到底)。
- 口语点缀按段配给(每段恰好一个"接地气"词),温度像用滴管加的。
+- 刻意口语补丁:抽象表述修掉后换上讨巧的口语(「延续之前积累的信息」改成「照样用得上」),修补本身读着就刻意;落点是最平实的陈述(之前积累的信息也还在)。
- 段段短句收尾的"留白"节拍器(人写的段落有时就平平地结束)。
- 判断句系列:连续段落都用「……的是X」承重(把流程定下来的是创始团队。真正把成本压下来的是第二代产线。),同一句型当引擎。
- 电报体编年:连续分句各带一个年份(1997年……。1998年……;2000年……,2001年……),叙事被压成 changelog;改用相对时间(三年后、世纪末)并让重要事件独占一句。