# aigc-reduce routes around the sampler, and bans the two moves that would defeat it > A Claude Code skill for lowering AIGC detection rates on academic papers, built from prompt rules and a nine-dimension scanner. Its most interesting constraints are prohibitions: no full rewrite, no colloquial register, no invented facts, and substitutions that never pass through token sampling. **xiaofenggan01/aigc-reduce** — 降低学术论文 AIGC 查重率的 Claude Code Skill | A Claude Code skill for reducing AIGC detection rates in academic papers - Repository: https://github.com/xiaofenggan01/aigc-reduce - Stars: 610 · Forks: 23 - Language: Python - License: MIT - Published: 2026-09-20 · Updated: 2026-09-20 · Language: en - Canonical page: https://hysenlabs.com/projects/xiaofenggan01-aigc-reduce ## The design routes around token sampling entirely There is no model in this repository. The primary language is Python because of one scanner script; everything else is Markdown rules, a skill file and reference documents. That makes the design legible in a way a trained system is not. Three of the four stated rules exist to keep text away from a language model while it is being edited. The first forbids a full rewrite, on the stated grounds that rewriting machine text with a machine stacks fingerprints, with a measurement attributed to one commercial detector showing the rate climbing to 100 percent after a rewrite. The third requires every replacement to be deterministic, changing a small span at a time and explicitly not passing through token sampling. That is the whole mechanism. If the sampler is what leaves the trace, then a rule set that performs fixed substitutions produces text the sampler did not produce. The cost is that every edit is a lookup, which is why the repository carries a replacement table, a list of high-frequency Chinese words used by machine writers, and a separate negative list of colloquial forms to avoid. The first pass is subtractive and specifies its own budget: word-level replacements, sentence-level restructuring and paragraph-level adjustment, each with a percentage band, adding up to a required modification rate above 40 percent. The second pass is additive, injecting written-academic features such as sentence-length variation against a target coefficient of variation near 0.45. The design is internally coherent. It is also, by construction, a description of how to look like something other than what wrote it, which is why the constraints in the next section matter more than the substitutions. ## Preserving register outranks the modification target The page states its core position in a single line: reducing the detection rate is not the same as colloquialising. The argument is that the default direction of de-AI-ification drifts toward informal language, because informal language does fool statistical detectors, but an academic paper still has to survive a human reader. So keeping academic written register is set as a hard floor, with a stated priority above the modification-rate goal. Rule two carries the same conflict resolution in writing: the 40 percent target must be reached, but only through structural rewriting such as clause order, sentence patterns, active and passive voice and splitting long sentences, and only by removing template sentences. Colloquialising, emotional language and meaningless word swaps are forbidden as ways of hitting the number. And when reaching 40 percent conflicts with preserving register, register wins. A numeric target subordinated to a qualitative floor is unusual in a tool of this kind, and it is the load-bearing decision. It also means the tool can legitimately fail to hit its own number, which is what the scanner's last two dimensions exist to catch. Two further prohibitions are worth naming because they are about honesty rather than style. The text must not acquire facts, data or references that were not in the original, and the second pass is restricted to completing observations and parameters already present. A tool that rewrites a paper is in a position to invent a citation by accident, and the rule closes that off explicitly. ## Two of the scanner's nine dimensions check the tool's own mistakes The repository ships a scanner that measures nine dimensions of a text file, and the mapping to detectors is the part to read as claims rather than as established fact: ```bash python3 scripts/aigc_scan.py your_paper.txt ``` Seven dimensions are presented as machine-writing traces: template-sentence density, sentence-length variation, paragraph symmetry, nested numbering, colon lists, passive voice and punctuation regularity. Each is attributed to a named product's technique, from semantic fingerprints at one provider through to a stated core algorithm at another and sentence-feature analysis at a third, with one dimension attributed to a transformer-based semantic analysis. The last two are different in kind. Colloquial register and dash density, with a ceiling of one dash per paragraph, are labelled as over-reduction warnings rather than machine-trace indicators. The page says so explicitly: they are gates against the tool having broken the register, and a hit points you at a reference document for restoring written academic style. That is the design detail worth carrying away. Two of nine checks exist to catch the failure mode the tool itself introduces, and they are weighted the same as the seven trace detectors rather than being a footnote. It also explains why rule four in the protocol, the audit pass, asks eleven questions about why a passage still reads as machine-written and then adds three register guardrails on top, including whether anything was invented to raise the modification rate. ## The false-positive numbers contradict each other on one page The strongest argument the page makes is about measurement rather than about detection. It states that results differ across platforms for the same paper, giving 35 percent on one domestic detector against 12 percent on another, and describes AIGC detectors as fundamentally limited. The supporting figures are given as false-positive rates: 6.5 percent for high-quality fluent papers against 1.8 percent for papers with informal flaws, 8 to 12 percent for terminology-dense papers above a stated term-density threshold, and a claimed 60 percent drop in detector accuracy after preference optimisation, used as evidence that the detectors rely on shallow features. The problem is that the reference tables on the same page give different numbers. One domestic product is listed at 1.2 percent false positives, which cannot be reconciled with the 6.5 percent figure quoted two sections earlier without an explanation, and the international table shows one product with a 16 percent false-positive rate alongside a stated accuracy range of 90 to 99 percent, with another product's false-positive rate listed as not disclosed. So the page's own reference data is uneven, and a reader who quotes a figure from it should quote which table it came from. The cross-platform inconsistency claim, by contrast, needs no numbers from this page to stand up. The consequence for anyone using the tool is direct. If the same text can score 35 on one platform and 12 on another, then a target modification rate is a target against an instrument whose reading depends on which instrument you pick. ## The structure listing omits the script it documents The page gives a file tree for the repository and then, a few lines later, tells you to run a script that the tree does not contain. The tree shows a skill file, the readme, an agents directory holding one metadata file for Codex, five reference documents covering the positive academic style standard, protected spans, replacement tables, AI trace patterns and detector principles, and an evaluations directory with a single benchmark file. That last line is itself cut off partway through its comment, so the benchmark's description never finishes. What is missing is the scripts directory and the tests directory, both of which are present in the repository and neither of which appears in the tree. The scanner that the page documents as the automated checking tool lives in the omitted scripts directory. That is a documentation gap rather than a design flaw, and it is a common one in a repository whose readme was written for humans skimming a skill rather than for someone about to run something. Still, the tree reads as exhaustive and is not. The agents directory is the other piece of structure worth explaining. One repository serves two clients: the skill is recognised through frontmatter in the skill file, and Codex additionally reads user-interface metadata from a file in that directory, which Claude Code is told it can ignore. The install path differs by client too, a per-user skills directory for Claude Code and a different discovery directory for Codex, with an environment-variable path for setups that use one. ## It triggers on keywords, and it names its sources Installation is a clone into a skills directory, and the trigger is keyword-based rather than always-on: once installed, mentioning phrases about reducing the AI taste, lowering the AIGC rate or removing AI detection causes the skill to trigger. That is a wide trigger surface for a tool that rewrites text, and it is worth knowing that installing it means those phrases will be acted on in future sessions rather than requiring an explicit invocation. The acknowledgements section is unusually specific about where the method came from, listing a research report on AIGC detection methods dated May 2026, a Wikipedia project page on signs of AI writing, an existing humanizer skill, prompt engineering validated on a forum, a sentence-transformation methodology from a social media account, a prompt published on a developer community, three open-source projects on the same problem, a fourth repository credited specifically for the positive-style contract, protected spans and bidirectional benchmark design, and feedback from users of one commercial detector. The maintenance state is small: no published releases, an MIT licence, and a last push dated 2026-07-08, so the rule set and the reference tables are pinned to whatever that commit contains rather than to a version anyone can name. ## Conclusion Treat this as an artefact worth reading for its false-positive argument rather than as a tool to adopt. Its central claim is a measurement claim, that the same paper scores 35 percent on one platform and 12 on another, which is a problem with the detectors rather than with the writer. Three limits before anything else. Do not use it to disguise machine-written work; that is the case its own rules are structured to refuse, and misrepresenting authorship is not something a writing skill can launder. Its accuracy and false-positive figures are the project's own, and two of them contradict each other inside the same page. And detectors change, so any fixed modification target is a moving one. ## FAQ ### What is aigc-reduce? An MIT Claude Code and Codex skill for lowering AIGC detection rates on academic papers. It contains no model: it is a skill file, five reference documents, an evaluation benchmark and one Python scanner, with a required modification rate above 40 percent reached through structural rewriting only. ### What does aigc-reduce scan for? Nine dimensions in one script: seven for machine-writing traces such as template-sentence density, sentence-length variation, paragraph symmetry, nested numbering, colon lists, passive voice and punctuation regularity, plus two for over-reduction, being colloquial register and dash density above one per paragraph. ### How do I install the aigc-reduce skill? Clone the repository into the skills directory for your client. For Claude Code that is a per-user Claude skills folder; for Codex it is a Codex discovery directory, or a `CODEX_HOME/skills` path if your setup uses that variable. Codex reads UI metadata from an agents file that Claude Code can ignore. ### Does aigc-reduce make the writing less formal? No, and that is its stated core position: reducing the detection rate is not the same as colloquialising. Keeping academic written register is a hard floor with priority above the modification-rate target, and colloquialising, emotional language and meaningless word swaps are forbidden as ways of reaching the number. ### What detector accuracy figures does aigc-reduce report? Tables for three domestic and three international products, with the strongest claim being that the same paper scores 35 percent on one platform and 12 on another. The page's own figures are uneven: one product is listed at 1.2 percent false positives while an earlier paragraph quotes 6.5 percent for fluent papers. ### What rules does aigc-reduce forbid? Four. No full rewrite by a model, since that stacks fingerprints. The 40 percent modification target may only be met by structural rewriting. Replacements must be deterministic and must not pass through token sampling. And the rewrite must never introduce facts, data or references absent from the original, nor network slang, emotional words or heavy dash use. ## Sources - [Issues](https://github.com/xiaofenggan01/aigc-reduce/issues) - [License: MIT](https://github.com/xiaofenggan01/aigc-reduce/blob/main/LICENSE) - [README](https://github.com/xiaofenggan01/aigc-reduce/blob/main/README.md) - [xiaofenggan01/aigc-reduce on GitHub](https://github.com/xiaofenggan01/aigc-reduce) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/xiaofenggan01-aigc-reduce