version: "1.0.1" name: humanly description: "Remove AI writing patterns from text. Three modes: (1) prewrite — read before writing to internalize anti-AI-slop rules, (2) rewrite — full post-writing audit with severity tiers and two-pass rewrite, (3) review — flag-only audit without rewriting. Use when: editing drafts for AI tells, cleaning up AI writing, auditing copy, making text sound human, 去 AI 味, 潤稿, 檢查 AI 味. Also use as a prewrite reference before composing any public-facing text."

Humanly — AI Writing Pattern Removal

Remove signs of AI-generated writing. Based on Wikipedia's "Signs of AI writing" and the avoid-ai-writing skill.

Core principle: If you wouldn't say it, don't write it. Write like a smart friend talking.

Order of operations: protect the facts, then strip the AI patterns, then add voice. In that order. A sentence with a point of view that broke a fact is worse than a boring sentence that is correct. And never invent a number, a source, or the author's memories to fill a hole you just made — leave a placeholder and hand it back. See references/protected-list.md.

Mode Detection

Prewrite Mode

When invoked before writing:

1. Detect target language from context
2. Read the corresponding generated prewrite file (path relative to this SKILL.md):

• 中文 → references/generated/prewrite-zh.md
• English → references/generated/prewrite-en.md

1. Internalize the principles and examples, then proceed with the writing task

The prewrite file already bundles the core principles, the "examples teach shape, not license to fabricate" rule, the highest-frequency patterns with before/after examples, the Tier 1 word table (plus banned sentence patterns and the Taiwan localization rules for zh), and a one-line index of every pattern. Do NOT additionally load the full patterns file — it is for rewrite / review mode.

The zh bundle carries the mainland→Taiwan vocabulary and full-width punctuation rules on purpose: the model's Chinese training data is mostly simplified, so it reaches for 視頻 / 質量 / half-width commas while writing. Catching that only in rewrite mode is catching it too late.

The zh prewrite bundle is zh-TW. There is deliberately no mainland variant — this skill's Chinese layer targets Taiwan, and a second generated bundle would be a whole parallel artifact for an audience the catalog isn't written for. If you are composing for mainland readers: don't load the zh prewrite bundle. Write first, then run rewrite mode and tell it to skip the Taiwan layer.

Rewrite / Review Mode Pipeline

Step 1: Detect Language

Determine the primary language of the input text.

Step 2: Load References

Read these files (paths relative to this SKILL.md):

Step 3: Detect Context Profile

Auto-detect from content cues (see context-profiles.md), or accept user hint:

• social / social-zh / blog / technical-blog / investor-email / docs / support-email / casual

Apply the tolerance matrix — some rules are relaxed or skipped per profile.

Step 4: Lock the Protected List

Before touching a word, circle what may not move: prices and numbers, proper nouns, links and anchor text, real names and directly quoted speech, commitments (refund policy, warranty, disclaimers, legal wording), and code — commands, paths, API routes, version strings. Count them — you will verify the same count in Step 9.

Quotation marks alone are not protection: scare quotes and emphasis quotes stay editable. The test is attribution — can you name who said it?

Full definitions, the false-positive table, and the never-invent rule: references/protected-list.md.

Step 5: First Pass — Scan by Severity

P0 — Credibility killers (fix immediately):

• Cutoff disclaimers ("As of my last update")
• Chatbot artifacts ("I hope this helps!", "Great question!")
• Vague attributions without sources ("Experts believe")
• Significance inflation on routine events
• Hallucinated citations — decimal-precise studies, misattributed quotes. Mark [source unverified] / 〔需查證來源〕, keep the sentence verbatim, never verify or invent
• AI tool residue — utm_source=chatgpt.com, turn0search0, citeturn. Grep for these; it is the one class you can catch mechanically
• Unfilled template placeholders — [Product Name], [insert case study]. Flag them, never fill them in. Not merge tags in an actual template, and not deliberate anonymization — see the boundary on the catalog entry

P1 — Obvious AI smell (fix before output):

• Tier 1 word violations
• Template phrases and slot-fill constructions
• "Let's" transition openers
• Synonym cycling within a paragraph
• Formulaic openings
• Bold overuse, em dash frequency

P2 — Stylistic polish (fix when possible):

• Generic conclusions
• Rule of three
• Uniform paragraph length
• Copula avoidance
• Transition phrases

Step 6: Cross-Language Checklist

Run through these checks regardless of language:

• Three consecutive sentences same length? Break one up
• Three consecutive sentences all short, hammering for drama? Merge them back — that is the same metronome, just faster
• The piece takes no position at all ("both have their merits", "it depends on the person")? Ask the author which one they picked and why. Never pick for them
• Opens with an era-hat ("In today's rapidly evolving landscape")? Delete the first paragraph and ask whether the piece lost any information. If it didn't, the paragraph was warm-up. If it did, keep it — some era-openings are the argument
• Paragraph ends with a tidy one-liner? Vary the ending
• Em dash before a reveal? Remove it
• Explaining a metaphor? Trust the reader
• Conjunctive adverbs (Additionally, However)? Consider removing
• Rule of three? Use two items or four
• Symmetrical slogans ("Not X, but Y")? Just say Y
• Contrast adds no new information (surface/deeper, "the real problem is")? Cut the frame, state the point with evidence
• Ends with a life lesson or quotable line? Delete or replace with a concrete fact
• More than 1 quoted term? Keep only the most essential one. This means emphasis and scare quotes — attributed speech is exempt and never gets compressed away, however many quotations the piece carries (see protected-list.md §4)
• Announcement filler ("You won't believe...")? Just say the content

Step 7: Quality Scoring

Score on 5 dimensions (1-10 each, total 50):

Thresholds: 45-50 excellent, 35-44 good, below 35 needs another pass.

Step 8: Second Pass Audit

Re-read the rewritten version:

1. Identify any remaining AI tells that survived the first pass
2. Check for recycled transitions, lingering inflation, copula swaps
3. Check what you introduced: fake-candor hooks ("honestly,"「說真的」), staccato drama, manufactured aphorisms, an invented anecdote or change of heart. Removing slop by adding a different slop is the most common failure of this pipeline
4. Fix and note what changed
5. If score was below 35, repeat from Step 5

Step 9: Fidelity Read-Back

Not optional. Run it even in automated pipelines.

1. Re-count the items locked in Step 4 and confirm each one survives verbatim
2. Confirm no fact, number or source appears that was not in the source text
3. Confirm the author's position did not flip or soften
4. Confirm the register held — a notice still reads as a notice

Any check fails: fix it, don't ship it.

Step 10: Output

Rewrite mode — return 4 sections:

1. Issues found: every AI-ism identified, quoted, with severity (P0/P1/P2)
2. Rewritten version: clean version
3. What changed: brief summary of major edits
4. Second-pass audit: surviving tells fixed, or "clean"

Review mode — return 2 sections:

1. Issues found: grouped by severity (P0/P1/P2)
2. Assessment: which flags are clear problems vs. judgment calls

Markers, in the target language. When cutting the filler leaves a hole only the author can fill, mark it in place and move on:

A rewrite that comes back mostly markers is not a failure. For a draft that was all air, it is the correct result — the ball goes back to the author.

Self-Reference Escape Hatch

When writing about AI patterns (blog posts, tutorials, documentation): quoted examples, code blocks, and text explicitly marked as illustrative are exempt from flagging. Only flag patterns in the author's own prose.

Maintenance

Seven source files, two generated files (built from four of them), one build script:

To change anything: edit the source file, then run

python3 skills/marketer/humanly/scripts/build-prewrite.py

--check verifies the generated files are current (the script also fails on missing summary lines or non-contiguous numbering). Each source file's header comment says exactly where new content belongs.

Reference

Based on Wikipedia:Signs of AI writing and avoid-ai-writing v3.3.0 (MIT License).

The fidelity layer (protected-list.md), the Taiwan localization layer, zh patterns #41–#50, and en patterns #39–#41 are adapted from Raymondhou0917/speak-human-tw (MIT License).

Sources: blader/humanizer, brandonwise/humanizer, hardikpandya/stop-slop, op7418/Humanizer-zh