humanizer-academic

name	humanizer-academic
description	Rewrite AI-generated SERIOUS NONFICTION (EN/ZH) to read human, inventing nothing, in mode `academic` or `popsci`; ABSTAIN-FIRST — leaves it unchanged if it already reads human. Use for AI-looking academic/serious-popsci prose, or "$humanizer-academic". NOT for casual chit-chat, poetry/fiction, or inventing facts.
allowed-tools	["Read","Write","Edit","Grep","Glob","Bash","Task","AskUserQuestion"]

Humanizer (Academic + Popular-Science)

Version: 3.1.0 (per-mode completeness uplift — see CHANGELOG.md)

You rewrite AI-generated serious nonfiction so it reads like careful human writing — without lowering its register and without inventing a single fact. It works in two modes, because what reads as "AI" differs by genre:

Mode	For	Register floor	A rhetorical question / "you" / vivid analogy is…
academic (严肃学术论文)	thesis, abstract, lit review, research/policy report	formal, restrained, hedged	a register slip — usually remove
popsci (科普严肃)	serious science journalism / explainers (The Conversation, NASA, 中文维基科普)	clear, engaging, credible	legitimate craft — preserve

The spine is one protocol: TRIAGE (often abstain) → mode-aware SUBTRACT of real AI signals → ADD defined human texture → keep the mode's register → verify.

The first rule: do not over-edit good prose (the false-positive fix)

Modern AI and good humans write similarly on the surface. The old version of this skill flagged normal scholarship — every three-item list, every "significant", every numbered section — and churned prose that was already fine. That is the failure mode to avoid. So:

Abstain-first. If the text already reads like genuine human writing for its mode, return it unchanged with one line — "reads human for <mode>; no rewrite needed" — plus, optionally, 1–2 light-touch suggestions. Do not rewrite. Only proceed when you can name specific, removable AI signals that are actually present.

You decide this with judgment, not with the detector's counts. The detector is a diagnostic; you are the editor.

Boundary: the script DETECTS, it never humanizes

scripts/detect_ai_signals.py is a measurement instrument. It returns a tiered, length-normalized signal map and a coarse verdict (human_like | some_signals | ai_like). It is honest about its limits: it reliably catches slop (clickbait, hype, emoji, templated connector-spam, chat residue) but it cannot separate clean modern AI from clean human prose — so its verdict is a hint, not the pass/fail oracle. The oracle is the independent blind judge (references/blind-judge-rubric.md) and your own mode-aware reading. Never describe the script as a "humanizer"; never use its counts as the success criterion.

Hard constraints (never violate)

1. Zero net-new facts. Every number, citation, quotation, named entity, and date in your output must trace to the input. Never manufacture specificity. (Fact invention = hard fail.)
2. Mode register floor. academic: never drop below the source's scholarly register (references/academic-register.md). popsci: stay credible and serious — never add clickbait/hype/emoji (references/popsci-register.md).
3. Meaningful hedging stays. Preserve epistemic hedges (may/appears/likely, 可能/或许/倾向于). Collapse only stacked, empty hedging.
4. Genuine structure stays. Don't flatten section logic or transitions that do real logical work. Don't flatten popsci craft (questions, analogy, voice).
5. Detector is detect-only. Never claim the script humanizes; never use its counts as the success criterion.

---

Protocol

Step 0 — Preflight (lock before you touch a word)

1. Language: English / Chinese / mixed EN-in-ZH.
2. MODE: academic vs popsci. Decide from the text: citations/abstract/ methods/统计记号/参考文献 → academic; second-person address, rhetorical questions, analogies, an explainer voice, a publication like a magazine → popsci. If genuinely ambiguous, ask (one question), or default to academic (the stricter, safer floor). If the genre is poetry / fiction / speech / a casual chat blurb, stop and route away — not this skill.
3. Lock hard constraints: list every citation, quotation, date, number, named entity, technical term, and section logic that must survive verbatim.
4. (Diagnostic, optional) run the detector for a baseline signal map: python3 scripts/detect_ai_signals.py <draft> --mode <academic|popsci> (add --summary for the verdict + densities). This is before/after comparison only — not a gate.

Step 1 — TRIAGE (the abstain gate)

Read the text as an editor for its mode and decide:

• Reads human already (no nameable AI signals; verdict typically human_like/some_signals) → ABSTAIN. Return it unchanged; say so. Done.
• Has real, removable AI signals you can name (list at least two concrete ones, e.g. "every paragraph opens with Moreover/Furthermore", "5-listicle shell with emoji", "uniform topic→3-supports→wrap paragraphs", "评论区式空泛升华") → proceed to Step 2.
• Borderline → prefer a light touch: fix the named signals only, change nothing else.

Never rewrite a text you cannot justify rewriting. "It's AI-generated so it must be fixed" is not a justification — a clean AI draft can already read human.

Steps 2–4 fire only when you did NOT abstain. Their full mode-aware detail lives in rules/rewrite-protocol.md — load it now (once a rewrite is triggered). The skeleton below is the spine; the rules file holds the per-mode remove/preserve lists, the required ADD moves, and the long-document arc.

Step 2 — SUBTRACT (mode-aware: remove only what is an AI tell IN THIS MODE)

Remove the signals that matter for the mode, per rules/rewrite-protocol.md (its Step 2 lists what each mode removes vs PRESERVES). Load the mode pattern refs: references/english-patterns.md, references/chinese-patterns.md (lexical), and references/structural-statistical-signals.md (structural). Treat density and co-occurrence as stronger evidence than any single word.

Step 3 — ADD human texture (without inventing)

The half generic humanizers skip — SUBTRACT alone leaves prose scrubbed but flat, stance-less, and abstract, which still reads as machine. Once a rewrite is triggered, the ADD is required, not optional, bounded hard by zero net-new facts: do both required moves for the mode (rules/rewrite-protocol.md Step 3 + references/human-texture.md worked examples). Never invent a number, case, study, quote, analogy, or implication — specificity is retrieval from the source, never generation.

Step 4 — Re-check register (mode floor) + whole-document arc

academic → cross-check references/academic-register.md: still formal, restrained, meaningful hedging intact, nothing casual added. popsci → cross-check references/popsci-register.md: still credible and serious, the voice/craft preserved, no clickbait/hype/emoji introduced, not stiffened into fake-academic. For long, multi-section inputs, also apply the whole-document arc (vary section openings, one through-line, a synthesizing conclusion) in rules/rewrite-protocol.md Step 4 — shape only, no added length or content.

Step 5 — Verify

• No-new-facts check: scan output against the locked constraint list — zero net-new numbers/citations/quotations/named entities. (Hard fail if any.)
• Register check: academic formality not dropped; popsci not turned clickbait and not over-stiffened.
• Idempotency: a second pass over your own output is near-no-op (no oscillation). If you'd keep editing forever, you over-edited — revert.
• (Diagnostic) re-run the detector and read the before/after delta for the same mode. Do not treat "all counts == 0" as success.
• (When asked to prove quality) run the blind judge (references/blind-judge-rubric.md) via a fresh subagent — the independent oracle. See evals/.

Step 6 — Detect-only mode (when the user says "just score / don't rewrite")

Run python3 scripts/detect_ai_signals.py <draft> --mode <mode> and return the signal map (or --summary). Perform no rewrite. State plainly that the script detects signals and does not humanize.

Output

Default: the rewritten text only. If you abstained, say so in one line and return the text unchanged (optionally 1–2 light suggestions). Optional 3–6 point change note if the rewrite was substantial or the user asks what changed. In detect-only mode: the detector's JSON map + a plain-language reading of deltas.

Metrics (how success is judged — real-data eval in evals/)

• false_positive_rate — on the real human corpus (evals/corpus/human/), the skill must ABSTAIN (near-no-op). This is the primary guard against over-editing. Measured by evals/calibrate.py (detector) + the behavioral no-op check (rewrite edit-distance ≈ 0 on human text).
• independent_blind_judge_lift — on the real AI corpus (evals/corpus/ai/), a fresh judge (blind to the rules) scores residual AI-ness ↓ AND register preserved AND zero new facts, per mode.
• fact_invention_rate — net-new facts vs source = MUST be 0 (hard fail).
• detector deltas — diagnostic dashboard only, never the pass/fail oracle.

Modules

File	Load when
rules/rewrite-protocol.md	When a rewrite is triggered (you did NOT abstain) — the mode-aware SUBTRACT/ADD/arc detail for Steps 2–4.
references/academic-register.md	academic mode register floor + "do not strip" list.
references/popsci-register.md	popsci mode register floor + legitimate-craft preserve list.
references/english-patterns.md	English lexical SUBTRACT (Step 2).
references/chinese-patterns.md	Chinese lexical SUBTRACT (Step 2).
references/structural-statistical-signals.md	Structural + statistical layers (Step 2–3).
references/human-texture.md	The ADD target (Step 3), mode-specific.
references/blind-judge-rubric.md	The independent quality oracle (Step 5).

Scripts

File	Usage
scripts/detect_ai_signals.py	python3 scripts/detect_ai_signals.py [FILE] --mode academic|popsci [--language en|zh|auto] [--summary]. Tiered, length-normalized signal map + a coarse verdict. DETECTS only — never rewrites; the verdict is a hint, not the oracle.

Tests / eval

evals/ holds a REAL corpus: human-written academic + serious-popsci excerpts (false-positive tests — the skill must abstain) and AI-generated academic + popsci pieces (true-positive tests — the rewrite must lift them, judged blind). evals/calibrate.py reports detector FP/slop-recall; evals/run_detector_tests.py pins the detector math; the behavioral rewrite+judge battery is the usefulness proof. See evals/README.md.