humanize

humanize — de-AI writing skills (prose + technical docs)

Two Claude/Codex/Cursor skills that make writing read less like AI, grounded in research rather than a vocabulary blocklist:

• humanize-prose — fiction, essays, blog, narrative non-fiction. Two layers: a surface pass (Wikipedia "Signs of AI writing") and a narrative pass (the StoryScope paper's discourse-level levers). Ships a working trained detector (reproduces StoryScope at 94.9% macro-F1) to score before/after.
• humanize-tech-docs — PRDs, TDDs, design docs, RFCs. Surface pass + structural tells (vague buzzword requirements, unowned passive voice, missing specifics, no rejected alternatives). Backed by SpecScope, a novel pilot detector we built (see below).

Install

bash tools/humanize/install.sh            # symlinks both skills into ~/.claude, ~/.codex, ~/.cursor
bash tools/humanize/setup-detector.sh     # one-time: builds the prose detector (optional; ~6 min)

install.sh symlinks (doesn't copy), so the repo stays canonical — edit here, agents pick it up. Everything except the prose Option-2 detector works with zero setup; the detector's large trained artifacts are gitignored and rebuilt from the public StoryScope release by setup-detector.sh.

Layout

tools/humanize/
  install.sh              symlink both skills into agent skill dirs
  setup-detector.sh       build the prose detector artifacts (venv + StoryScope clone + train)
  requirements.txt        detector Python deps
  skills/
    humanize-prose/       SKILL.md · references/ · detector/ (scripts+taxonomy; artifacts gitignored)
    humanize-tech-docs/   SKILL.md · references/ · detector-pilot/ (SpecScope: code + corpus + FINDINGS)
  research/
    RESUME.md             full project arc, current state, and exact next steps — READ THIS to resume
    landscape.md          GitHub survey of humanizer skills + the head-to-head detector results
    comparisons/          the 4-way skill comparison (AI original + 3 humanizer rewrites + features)

Origin & evidence base

• StoryScope (arXiv 2604.03136, COLM 2026): discourse-level narrative features detect AI fiction at 93.2% macro-F1 and survive surface rewrites. We reproduced its classifier from the authors' released corpus (94.9% / 70.1%), which powers humanize-prose's detector.
• SpecScope (ours, humanize-tech-docs/detector-pilot/): the StoryScope method applied to specs — no prior art existed. 117 real specs (Rust RFCs, K8s KEPs, React RFCs) vs Claude mirrors, 27 spec-discourse features, XGBoost. Pooled 96.6% macro-F1; 92.5% mean under leave-one-domain-out. Robust signals: humans cite the real world (URLs/issue refs; AI≈0), are lumpy (uneven sections/sentence length), and hedge / leave TBDs. Full write-up in detector-pilot/FINDINGS.md.

See research/RESUME.md to pick the work back up.