de-ai plugins

by skymanbp

Scientific paper writing, source-traced review, and de-AI rewriting for top-tier journals (ApJ/MNRAS/PRD/JCAP-class) and NSF/NIH proposals. Ships 12 skills and 40 tools under one typed standard: scientific-integrity blockers that cannot be waived as style preferences, a zero-target L0 lexicon, ranked advisories, explicit measurement states so missing calibration never reads as clean, claim-first rewriting gated on hard scientific-fidelity eligibility, a mechanical length budget with a machine-built removal map and a required-shrink gate, corpus-calibrated salience-hierarchy, domain-register (with an exhaustive zero-hit vocabulary audit), collocation, and discourse-texture (given/new cohesion, epistemic hedging) axes, each judged against the field's own per-section distribution at the unit it was calibrated at, a residue axis that gates the trace an edit leaves, A-R review with narrative-spine and adversarial-escalation dimensions, registry-verified bibliographies (CrossRef, DataCite, arXiv), pixel-measured figure review, and isolated multi-reviewer orchestration. Learned scores are field-similarity triage, never an authorship verdict. See CHANGELOG.md for per-version history.

by XRenSiu

Humanize — make AI-written technical prose read like a competent human professional wrote it, at the level where readers actually get lost: discourse, not vocabulary. Three skills. `/humanize` rewrites a draft (zh/en) without changing a single fact: it targets the mechanisms behind 'every word is clear but I can't follow' — given-new inversion, broken topic strings, stress-position leakage (participle / 以实现… tails), lists replacing argument, cross-section restatement, stance flattening, generic openings, summary closers — plus the solved lexical/syntax/format layer (Wikipedia 'Signs of AI writing' + 2026 density-era tics: mannered prose, invented jargon, 'not X, it's Y') and Chinese translationese (yage.ai four classes + 余光中 欧化 catalog, CCL-2023 corpus-backed). Three non-skippable gates: humanlint.py (23 surface metrics, zh/en, 0-100 AI-flavor index, calibrated on four fixtures: AI 60/63 vs human 2/0), factdiff.py (numbers / dates / identifiers / URLs / paths zero add-or-drop, plus certainty-drift warning per Belem 2026), and a context-isolated cold-reader agent that records per paragraph where a reader who never saw the brief loses the thread (expected / got / lost_at / told_not_shown) — never self-graded in the drafting context (self-refine amplifies self-bias); verify_coldread.py recomputes the reader's verdict from its five declared rules and enforces the 2-round cap mechanically (v0.2.0). Max 2 revision rounds, then deliver with the residual named. Genre decides whether structure may move (v0.3.0): proposal / design / adr / memo may be reorganised; report / readme / reference run with --keep-structure, where structdiff.py rejects any change to the heading-level sequence, list blocks and item counts, table shapes or code blocks — only the sentences inside prose paragraphs change; humanlint gained a matching `report` threshold tier. `/techdoc` writes a proposal / design doc / ADR / postmortem / memo from a brief in the shape senior engineers use (Google design docs, Oxide RFD, Amazon 6-pager, Nygard ADR, Shape Up, HashiCorp/Rust RFC, 阿里 安全生产): triggering incident with a number first, decision in one sentence, obvious alternatives killed with specific reasons, all consequences, rollout/rollback, open questions last — enforced as product order by verify_techdoc.py; facts only from brief/materials/domain knowledge, else [需核实]. `/voice-profile` distills the user's own 3–5 samples into an executable .humanize/voice.md (rhythm stats from humanlint, quoted sentence habits, personal 忌口表 with replacements, positive examples for completion-style few-shot — the technique with 20x measured style-match gain), gated by verify_voice.py. Ships rules/human-voice.md as an always-on contract pasteable into CLAUDE.md. Explicitly NOT for beating AI detectors (they train on humanizer output), NOT for injecting slang/emoji/typos as 'texture' (a new tell in technical register). All static_only tier: scripts smoke-tested on fixtures; behavioral delta on held-out drafts unverified.