[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-93933":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":16,"stars7d":16,"stars30d":16,"stars90d":16,"forks30d":16,"starsTrendScore":16,"compositeScore":17,"rankGlobal":10,"rankLanguage":10,"license":18,"archived":19,"fork":19,"defaultBranch":20,"hasWiki":21,"hasPages":19,"topics":22,"createdAt":10,"pushedAt":10,"updatedAt":30,"readmeContent":31,"aiSummary":32,"trendingCount":16,"starSnapshotCount":16,"syncStatus":33,"lastSyncTime":34,"discoverSource":35},93933,"humanizer-stack","NulightJens\u002Fhumanizer-stack","NulightJens","Two-pass pipeline for removing AI writing tells from outward-facing text: a surface pass plus a structural pass grounded in the StoryScope study. Packaged as Claude Code Skills.      Free community: skool.com\u002Fjens-ai-community-1306","https:\u002F\u002Fwww.skool.com\u002Fjens-ai-community-1306",null,"Python",131,9,104,1,0,43,"Other",false,"main",true,[23,24,25,26,27,28,29],"ai-detection","ai-writing","anthropic","claude-code","claude-skills","humanizer","writing-tools","2026-09-21 04:01:26","# humanizer-stack\n\nA two-pass pipeline for removing the signs of AI writing from outward-facing text,\npackaged as [Claude Code Skills](https:\u002F\u002Fcode.claude.com\u002Fdocs\u002Fen\u002Fskills).\n\nMost humanizers only fix words. That is the easy half, and it is the half that is\ndecaying fastest. This repo pairs a surface pass with a structural pass, because the\nresearch says structure is where the durable fingerprint lives.\n\n> **Free community.** I build tools like this in the open inside the\n> [Jens AI Community](https:\u002F\u002Fwww.skool.com\u002Fjens-ai-community-1306), a free Skool group\n> for putting AI to work in your business. If this repo is useful to you, come join us:\n> https:\u002F\u002Fwww.skool.com\u002Fjens-ai-community-1306\n\n## Why two passes\n\nThe [StoryScope study](https:\u002F\u002Fgithub.com\u002Fjenna-russell\u002Fstoryscope) (Russell et al.,\n2026) classified 61,608 stories from humans and five LLMs using only discourse-level\nfeatures, with every style feature withheld. It detected AI text at **93.2% F1**.\n\nThen the authors ran AI text through LAMP, a professional span-level rewriting system\nthat strips cliche, purple prose, and redundant exposition. Functionally, a very good\nsurface humanizer.\n\nDetection dropped **1.6 points**.\n\nMeanwhile the surface layer is eroding on its own. GPT 5.4 already cut its em-dash\nusage sharply, and fine-tuning drops stylistic detection from 97% to 3%. Word-level\ntells are a moving target. Structural tells require structural rewrites.\n\nSo: pass 1 fixes the words. Pass 2 fixes the shape. Run them in that order.\n\n## What is in here\n\n```\nskills\u002F\n  humanizer\u002F                     Pass 1: words and phrasing\n    SKILL.md\n    references\u002Fcopy-tells.md     Copy-specific tells (em dash, hype vocab, antithesis)\n  structural-humanizer\u002F          Pass 2: discourse structure\n    SKILL.md\n    references\u002F\n      storyscope-findings.md     The study distilled: 30 core features with rates\n      genre-calibration.md       Which audits apply per genre\n    scripts\u002Fstructural_scan.py   Deterministic scanner for grep-able structural tells\nscripts\u002F\n  copy_scan.py                   Deterministic scanner for mechanical copy tells\ndocs\u002F\n  PIPELINE.md                    How the passes chain, and what each one owns\n```\n\n### Pass 1: `humanizer`\n\nVocabulary, punctuation, and phrasing. Inflated symbolism, promotional language,\nsuperficial \"-ing\" analyses, vague attributions, em dash overuse, rule of three, AI\nvocabulary, negative parallelism. Built from Wikipedia's\n[Signs of AI writing](https:\u002F\u002Fen.wikipedia.org\u002Fwiki\u002FWikipedia:Signs_of_AI_writing).\n\nThe `copy-tells.md` reference adds the tells that show up specifically in public copy,\nranked by a 3.2M-post Reddit analysis of what people actually flag.\n\n### Pass 2: `structural-humanizer`\n\nSix audits run one at a time, because aspect-based checking found 95% of issues in the\nstudy's own pipeline against 68% for a single combined pass:\n\n1. **Theme explicitness.** AI states its lesson. Narrator explains the theme 77% of the\n   time against 52% for humans.\n2. **Structural tidiness.** Single-track, everything resolved. Humans digress and leave\n   threads open.\n3. **Emotion mode.** The largest gap in the study. AI performs emotion through the body\n   81% of the time against 38% for humans. Humans just name the feeling.\n4. **Reference specificity.** Humans name real things (47% against 24%). AI stays at\n   vague allusion.\n5. **Reader engagement.** Humans acknowledge the reader. AI writes as though no one is\n   watching.\n6. **Shape convergence.** Does this piece have the same skeleton as your last three?\n\n## The trap\n\nDo not trade one default for another. The study's deepest finding is convergence: all\nfive models occupy one tight region of structural space while humans are dispersed.\nRarity is the human signal.\n\nIf every piece now opens mid-scene, names three feelings, and ends unresolved, you have\nbuilt a new detectable cluster. Pick one or two interventions per piece, vary them\nacross pieces, and be able to say why this piece got this shape.\n\n## Install\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FNulightJens\u002Fhumanizer-stack.git\ncd humanizer-stack\n.\u002Finstall.sh\n```\n\nThis symlinks both skills into `~\u002F.claude\u002Fskills\u002F`, so updates land with a `git pull`.\nPass `--copy` if you would rather have independent copies than symlinks.\n\nTo install manually, copy `skills\u002Fhumanizer` and `skills\u002Fstructural-humanizer` into\n`~\u002F.claude\u002Fskills\u002F` (user-level) or `.claude\u002Fskills\u002F` (project-level).\n\n## Use\n\nIn Claude Code, the skills trigger on intent:\n\n```\nhumanize this post\nde-slop this lesson\nrun the structural pass on draft.md\n```\n\nRun the surface pass first, then the structural pass. `docs\u002FPIPELINE.md` covers the\norder and what each layer owns.\n\n### Scanners\n\nBoth scanners are deterministic and hook-friendly. They catch the pattern-matchable\nslice only, and neither replaces the judgment work in the skills.\n\n```bash\npython3 scripts\u002Fcopy_scan.py draft.md\npython3 skills\u002Fstructural-humanizer\u002Fscripts\u002Fstructural_scan.py draft.md\n\npython3 scripts\u002Fcopy_scan.py --json draft.md      # machine-readable\npython3 scripts\u002Fcopy_scan.py --strict draft.md    # exit 1 on any hit\ncat draft.md | python3 scripts\u002Fcopy_scan.py -     # stdin\n```\n\nMark a line `copy-ignore` to suppress an intentional usage.\n\n## Honest limits\n\n- **StoryScope studied roughly 5,000-word fiction.** Applying it to short nonfiction is\n  an inference, not a result the paper establishes. The subset that transfers most\n  cleanly is audits 1, 3, 4, and 6.\n- **Nothing here makes text undetectable, and that is not the goal.** The goal is\n  writing that reads as though a person with a specific point of view wrote it, because\n  a person did.\n- **The scanners catch maybe half.** Cadence, formulaic shape, and polished-but-empty\n  filler are only visible to a human reader.\n- **Audit 3 contradicts standard writing advice.** \"Show, don't tell\" is now a machine\n  signature. That is what the data says, and it is worth sitting with before applying.\n\n## Attribution\n\nBuilt on work by [@blader](https:\u002F\u002Fgithub.com\u002Fblader\u002Fhumanizer) (MIT),\n[jcarterjohnson](https:\u002F\u002Fgithub.com\u002Fjcarterjohnson\u002Fvibecoded-design-tells) (MIT), and\nWikipedia's WikiProject AI Cleanup (CC BY-SA 4.0). Grounded in Russell et al. 2026.\n\nFull breakdown with license obligations: [ATTRIBUTION.md](ATTRIBUTION.md).\n\n## License\n\nMIT for this repository's own work. Portions carry upstream terms, including CC BY-SA\n4.0 material with share-alike obligations. See [ATTRIBUTION.md](ATTRIBUTION.md) before\nredistributing.\n","这是一个专为消除AI生成文本痕迹而设计的双阶段重写工具链，包含表层语言优化和深层结构重构两个环节。核心功能基于StoryScope研究发现的结构性特征（如主题显性化、叙事节奏等），第一阶段修正词汇、标点与修辞习惯，第二阶段通过六类话语结构审计进行结构性重写；技术上以Claude Code Skills形式封装，采用确定性规则扫描与领域适配策略。适用于需要将AI初稿转化为自然、可信、面向公众传播文案的场景，如营销文案、品牌内容、教育材料及媒体发布前的合规性优化。",2,"2026-07-28 02:30:07","CREATED_QUERY"]