agentleFS
sl4dk1ysl4dk1y/human-writing/SKILL.md

human-writing

Skill1 starsChanged 7 days agoMITView on GitHub

Write and edit prose that does not read as AI-generated.

45 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use leader, storage-analyzer and 4 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • baizhine999Skill

    leader

    AI 扮演你的 Leader,用大厂管理思维审视你的方案、想法和工作,push 你成长。触发词:/leader, leader模式, 大厂模式, 帮我看看这个, 我有个方案, 我最近有点迷茫, 我不知道该怎么做。AI 是施压者,你是被 push 的那个人。

    Also in 3 projects that use this

  • xiaolinVVSkill

    storage-analyzer

    macOS / Windows 只读存储分析助手(自动识别系统)。扫描整机磁盘占用,找出 占空间大户,把每一项分成 🟢可自动清理 / 🟡需人工判断 / 🔴谨慎清理 三级并给出 可执行处置方案,生成排版精美、可折叠、命令可一键复制的交互式 HTML 报告,并可 起本地服务在网页上一键删除(移废纸篓/直接删)。扫描全程只读。务必在以下场景 使用:用户说"存储分析""磁盘满了""C盘/硬盘满了""空间不够""清理空间" "清理磁盘""占空间""哪些东西占地方""帮我看看存储""看一下电脑存储/空间" "存储空间""电脑空间不够""内存满了/不够/不足""看下内存/存储"(中文口语里 "内存"常指存储空间)"storage analysis""disk cleanup""清缓存""磁盘清理"; 或用户抱怨电脑没空间、想知道什么东西吃硬盘、想要清理建议时。注意:若用户明确 指运行内存/RAM(如"哪个进程吃内存""内存占用高"想看活动监视器),那是 RAM 不是存储,不属于本 skill。

    Also in 3 projects that use this

  • Wyx1105170322Skill

    neat-freak

    End-of-session knowledge cleanup with OCD-level rigor — reconciles project docs (CLAUDE.md, README.md, docs/) and agent memory against the code so nothing rots. 会话结束后对项目文档和记忆进行洁癖级审查与同步。MUST trigger when the user says: "sync up", "tidy up docs", "update memory", "clean up docs", "/sync", "/neat", "同步一下", "整理文档", "整理一下", "更新记忆", "梳理一下", "收尾", "这个阶段做完了", "新人能直接上手", or any phrase suggesting a dev milestone where knowledge needs reconciliation. Also trigger when the user reports stale docs, conflicting memories, or wants a clean handoff to teammates or other agents. Bare "整理" / "tidy" with prior dev context counts — do not under-trigger. Cross-platform: works on Claude Code, OpenAI Codex, OpenCode, and OpenClaw.

    Also in 3 projects that use this

  • NVIDIASkill

    deep-research

    Queue deep, multi-step research and analysis tasks to the DeepAgents worker. Supports execution depth presets, request-specific rubrics, SubAgent delegation, and RubricMiddleware cross-validation.

    Also in 3 projects that use this

  • VercelSkill

    find-skills

    Helps users discover and install agent skills when they ask questions like "how do I do X", "find a skill for X", "is there a skill that can...", or express interest in extending capabilities. This skill should be used when the user is looking for functionality that might exist as an installable skill.

    Also in 3 projects that use this

  • SentrySkill

    skill-creator

    Create new agent skills following the Agent Skills specification. Use when asked to "create a skill", "add a new skill", "write a skill", "make a skill", "build a skill", or scaffold a new skill with SKILL.md. Guides through requirements, writing, registration, and verification.

    Also in 3 projects that use this

The file

Show its rules and text

We split this skill into the rules it gives an agent

13 rules. 0 are common, 12 are distinctive, and 1 is very specific to this project.

  • 0

    Common

    Most of the 28 edited versions we read keep these.

  • 12

    Distinctive

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Not for code, logs, legal wording, or anything where the exact formulation is the point.

    line 20
  2. Distinctive:

    Every sentence you keep must tell the reader something they did not already have, either from earlier in the text or from the conversation …

    line 30
  3. Distinctive:

    If you have a sample of the author's own writing, read it first: sentence length, vocabulary, punctuation, openings.

    line 37
  4. Distinctive:

    Write from the fact. Each paragraph starts with what happened or what is known, not with how important it is.

    line 38
  5. Distinctive:

    Read it back and cut everything that adds no fact: announcements, run-ups, closers, statements of significance.

    line 39
  6. Distinctive:

    Check the rhythm aloud. Three sentences in a row of the same length means one of them gets shorter or longer, but not by formula.

    line 40
  7. Very specific:

    Run scripts/ai_tells.py (below) and look at what survived.

    line 41
  8. Distinctive:

    Check the facts: nothing added, nothing lost.

    line 42
  9. Distinctive:

    Every sentence adds a fact or a consequence.

    line 46
  10. Distinctive:

    No summary closer. "In conclusion, it is worth noting" restates the text in the same words.

    line 50
  11. Distinctive:

    Do not attach importance to objects.

    line 52
  12. Distinctive:

    Do not hide the author. First person and doubt are normal in personal writing.

    line 54
  13. Distinctive:

    Leave things imperfect. An opinion, an aside in brackets, a thought you did not finish, an admission that you do not know.

    line 55

“Common” is measured: kept by at least 6 in 10 of the 28 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Human writing

Machine-written text gives itself away through habits, not through passphrases. The rhythm is even, every thought is announced before it is made, importance is attached to things instead of to people, and nothing is left unfinished. Word lists rot with each model release. The habits last years. So the work goes from structure down to vocabulary, not the other way round.

When to use this

  • Writing something a person will read as human: a post, an email, an article, docs, a product page, a PR comment.
  • Editing text that already smells like a chatbot.
  • Reviewing someone else's text before it goes out.
  • Writing in any language. references/patterns.md holds the structural catalogue, references/evidence.md holds the measurements and sources, and references/lang-*.md holds per-language markers and typography.

Not for code, logs, legal wording, or anything where the exact formulation is the point. There the goal is the opposite: remove variation.

Language follows the audience, not the conversation. Chat replies stay in the user's language. An artifact meant for publication (a README, a SKILL.md, docs in a public repo) is written in English even when the discussion around it was in another language, with examples left in their own language where the examples are the subject.

The one idea

A model picks the wording that fits the widest range of readers and topics. A real author writes for one reader and one occasion, so their choices are uneven and specific. Everything below is a special case of one rule: the concrete over the average.

Two consequences follow.

Every sentence you keep must tell the reader something they did not already have, either from earlier in the text or from the conversation around it.

A tell counts in proportion to how rarely a careful writer would do it on purpose. One dash means nothing. A dash in every paragraph means a lot.

How to work

  1. Decide what the text is and who reads it. A personal post may say "I don't know". An onboarding doc may not.
  2. If you have a sample of the author's own writing, read it first: sentence length, vocabulary, punctuation, openings. The sample outranks every rule below, including the dash rule.
  3. Write from the fact. Each paragraph starts with what happened or what is known, not with how important it is.
  4. Read it back and cut everything that adds no fact: announcements, run-ups, closers, statements of significance.
  5. Check the rhythm aloud. Three sentences in a row of the same length means one of them gets shorter or longer, but not by formula.
  6. Run scripts/ai_tells.py (below) and look at what survived.
  7. Check the facts: nothing added, nothing lost. An invented detail is worse than a stock phrase.

Ten rules that carry most of the weight

  1. Every sentence adds a fact or a consequence. A line that only signals "something important is coming" is waste.
  2. Concrete over evaluative. Not "a revolutionary approach" but what was done and what it changed.
  3. Uneven rhythm. A short sentence after a long one works. An even paragraph of twenty-word sentences never does.
  4. No "not X but Y", no "it's not just a tool, it's a philosophy". The negative half invents an opponent, which makes the second half feel heavier without saying more.
  5. No summary closer. "In conclusion, it is worth noting" restates the text in the same words. The last line should carry a fact, not a promise.
  6. Say it instead of announcing it. "Let's dive in", "it's important to understand", "here's what you need to know" replace a thought with a trailer for the thought.
  7. Do not attach importance to objects. An event does not "underscore" anything, a town does not "symbolise" anything, an approach does not "play a key role". A person does that, and only if they said so.
  8. Formatting by need, not by rhythm. Bold, lists, headings and rules go where the text is unreadable without them. A four-item list almost always reads better as a paragraph.
  9. Do not hide the author. First person and doubt are normal in personal writing. Impersonal constructions belong in reference material.
  10. Leave things imperfect. An opinion, an aside in brackets, a thought you did not finish, an admission that you do not know. Completeness and smoothness are the machine signature.

Language packs

Structure transfers across languages. Markers and typography do not, and each pack lists what is specific.

PackCovers
references/lang-en.mdEnglish word lists, model-family favourites, the em dash history, British and American spelling
references/lang-ru.mdCalques, канцелярит, dangling gerunds, and the dash question, which is a real grammar mark and cannot simply be banned
references/lang-de.mdNominal style, -ung chains, "In der heutigen...", dash and quotation conventions
references/lang-es.md"No solo... sino también", "juega un papel crucial", gerundio riders, regional variation
references/lang-fr.md"Il est important de noter", "dans le monde d'aujourd'hui", nominalisation, French spacing an English writer breaks by accident

The Russian and English lists come from published field guides. The German, Spanish and French lists are compiled from observed model output: they are the same English habits translated, not corpus measurements. Say so if you lean on them in public.

Checking

python3 scripts/ai_tells.py text.md                # language detected
python3 scripts/ai_tells.py text.md --lang ru      # force a language
python3 scripts/ai_tells.py text.md --json         # numbers only

The script counts dash, bold, list and heading density, the spread of sentence lengths, matches against per-language phrase lists, closing formulas and staged openers. It strips code, URLs, quotations and > blocks before counting, because a guide about stock phrases would otherwise flag its own examples. The cost is that a quoted phrase never shows up in the report, which does not mean it is absent.

The thresholds come from published corpora, but they are heuristics, not verdicts. !! means look at it with your eyes, · means a soft signal that proves nothing alone. The script will happily flag headings and lists in any structured document, so check that mark against the genre instead of editing it away: an instruction needs headings, an email does not.

The two rhythm measures, mean sentence length and its spread, are reported but never judged below roughly 200 words and 12 sentences. At that size the spread varies as much inside one author as between authors, so a short text always shows · there and a !! only means something on a long one.

One limitation to expect: a document that lists markers as examples (like these language packs) reports its own lists as phrase hits, because only quotations and blockquotes are masked, not bare bullet points. On that kind of file, read the counts and ignore them.

Work order: fix !! from the top, then reread the whole thing. The phrase list is useful less for naming bad words than for printing line numbers, which gets you to the spot where a thought was replaced by an evaluation.

Example

Before:

In today's fast-paced world, note-taking plays a key role. It is not just a tool, it is a second brain. It is important to note that a good system provides clarity and fosters productivity.

After:

I kept a notebook for eleven years and then stopped. The reason is boring: I got tired of my own handwriting. What I remember from it is the household stuff. Buy stamps. Call the dentist back. The phone number of the man who fixed the boiler in 2016 and never picked up again.

The first version compresses to one sentence without losing anything, and that is the diagnosis. The second cannot be compressed, because it has details in it.

When not to touch it

  • The word or phrase sits in a quotation, a title, a name, or a text that discusses the phrase rather than uses it.
  • The author writes that way on purpose and it is their voice: repetition, dashes, exclamations, bureaucratic style used as a device.
  • The text was written before 30 November 2022. It cannot be machine-written by definition.
  • The genre needs evenness: an instruction, a regulation, a legal text, a reference entry. There the job is to remove ambiguity, not to add character.
  • AI detectors and a reviewer's gut feeling are wrong more often than people think. In one Japanese-corpus study, readers did barely better than chance while stylometry did much better. Judge on the pattern of evidence, not on one word.

What to keep when editing

  • An odd specific detail: a real address, someone's actual words, "the lawyer whose office was one floor above my dentist's".
  • Unresolved tension: "I think this is mostly fine, and it bothers me, and I cannot say why".
  • Period markers: slang, a meme, a joke that only lands in this year.
  • A choice the author can explain.
  • A real aside or self-correction: "(I wanted to write 'almost', but it really was certain)".

Sources

Numbers and links in references/evidence.md. Briefly: field guides are maintained by WikiProject AI Cleanup on English Wikipedia and separately on Russian Wikipedia, rates are measured by Pangram and Graphite on corpora in the millions of documents, vocabulary spikes are documented in Science Advances, and stylometry is covered in arXiv and PLOS ONE.

One finding worth carrying: by 2026 the em dash had stopped being a reliable tell, because newer models use it less than humans do. Individual words also differ per model family, so check the habit, not the word.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.