deslop
Skill69 starsChanged 11 days agoMITView on GitHub
Review or rewrite existing Chinese, English, or mixed-language prose to remove formulaic AI-slop patterns while preserving facts, intent, stance, register, and author voice. Use when the user explicitly asks to deslop, humanize, remove AI writing patterns or AI 味, 说人话, 别像模板/机器人, make text less AI-generated, or audit a draft for those problems. Do not trigger for generic polishing, proofreading, translation, summarization, fact-checking, or drafting from scratch unless deslop is explicitly requested as a step.
We split this skill into the rules it gives an agent
14 rules. 0 are common, 14 are distinctive, and 0 are very specific to this project.
0
Common
Most of the 107 edited versions we read keep these.
14
Distinctive
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 48
If a sentence contains only slop, keep an awkward source-faithful shell or the original sentence rather than inventing a concrete object or …
- Distinctive: line 56
Always read fidelity.md. Then load selectively:
- Distinctive: line 84
Find clusters, not tokens: Look for repeated formulaic structures, generic framing, performative transitions, flattened rhythm, and …
- Distinctive: line 87
Rewrite toward the scene: Prefer direct claims, concrete source-backed detail, natural sentence boundaries, and an appropriate amount of …
- Distinctive: line 93
Run the residual pass: Remove remaining high-confidence slop without sanding away legitimate voice or genre conventions.
- Distinctive: line 107
Consider no-op: If the source is already natural and fit for purpose, leave it alone.
- Distinctive: line 112
Do not invent facts, examples, citations, dates, measurements, product capabilities, feedback, emotions, opinions, or first-person …
- Distinctive: line 114
Do not turn vague attribution into direct fact, or uncertainty into certainty.
- Distinctive: line 115
Do not change the actor, owner, object, direction, completion state, causal relationship, comparison, or type of effect.
- Distinctive: line 117
Keep quotations, titles, proper names, technical terms, commands, paths, identifiers, logs, schemas, and literals exact unless the user …
- Distinctive: line 122
Treat unsupported claims as source content, not permission to fact-check.
- Distinctive: line 125
Never fetch or infer private voice samples.
- Distinctive: line 164
Do not flag isolated words without contextual evidence.
- Distinctive: line 166
If there are no material findings, say so and do not invent issues to fill the format.
“Common” is measured: kept by at least 6 in 10 of the 107 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.
Files
Deslop / 说人话
Edit existing prose so it sounds like a person making deliberate choices in a specific context, not a model performing “human” style. Preserve truth before style. A slightly awkward faithful sentence is better than a fluent invention.
This is an editorial skill, not an AI detector. Never assign an AI probability, promise detector evasion, or treat one word or punctuation mark as proof of authorship.
Route the request
Choose one delivery mode:
- Rewrite — default. Return one ready-to-use revision.
- Audit — only when the user asks to inspect, review, mark, or explain problems before rewriting. Report findings; do not silently rewrite.
- Embedded — when another workflow needs prose as an intermediate result. Return only the final text, without process notes. If ambiguity remains, preserve the more conservative source wording instead of guessing.
- File — only when the user explicitly asks to edit a file. Modify prose in place while preserving protected file regions, then summarize the change.
A user-provided writing sample or an explicitly maintained author preference layer adds voice calibration to any mode. It does not create a separate mode and never relaxes factual fidelity. Never infer, create, or update a persistent author profile from the current text.
Choose an edit scope:
- balanced — default. Reshape short text freely without deleting unique information. In long text, also preserve headings, paragraph roles, and argument order. Remove only demonstrably empty scaffolding.
- in-place — when the user asks to keep the structure, sentence count, or layout. Rewrite within sentences; do not delete, merge, or reorder them. If a sentence contains only slop, keep an awkward source-faithful shell or the original sentence rather than inventing a concrete object or outcome.
- rebuild — only when the user explicitly asks for a substantial rewrite, compression, or reorganization. Preserve the semantic ledger even when the shape changes.
Load only the needed references
Always read fidelity.md. Then load selectively:
- Chinese or Chinese-dominant mixed text: patterns-zh.md
- English or English-dominant mixed text: patterns-en.md
- Genre, register, or long-form decisions: scenes.md
- A supplied author sample, explicit voice-matching request, or approved author preference layer: voice-calibration.md
- Natural text, quotations, academic/legal/technical prose, or any uncertain match: boundary-cases.md
For mixed-language text, read both language pattern files but judge each phrase in its local linguistic and technical context.
Rewrite loop
- Read for meaning. Identify the document’s job, audience, register, and the author’s actual position.
- Build a silent semantic ledger. Record protected spans and relationships:
who did what, to what, when, with what status, evidence, strength, and effect.
For long text, also note each section’s job and what new claim, evidence, or
consequence it contributes. If a paragraph only paraphrases an earlier point
without adding a new fact, action, example, distinction, or consequence, mark
it as redundant — under
balancedscope it may be merged or removed; underin-placeit stays. - Find clusters, not tokens. Look for repeated formulaic structures, generic framing, performative transitions, flattened rhythm, and register mismatch. Do not rewrite merely because a watched word appears once.
- Rewrite toward the scene. Prefer direct claims, concrete source-backed detail, natural sentence boundaries, and an appropriate amount of irregularity. Do not manufacture personality.
- Run the fidelity pass. Compare every fact, number, name, attribution, relationship, modality, completion state, quotation, term, and protected file region against the source. Revert any unsupported change.
- Run the residual pass. Remove remaining high-confidence slop without sanding away legitimate voice or genre conventions.
- Check the ending. For long-form prose, test the final paragraphs: if deleting the last one or two paragraphs makes the piece stronger, end earlier. A final paragraph that only restates the thesis without a new consequence, action, or synthesis can be compressed. If the body never sustained a large scale (时代、文明、未来、世界、历史), do not let the ending escalate to it.
- Cold read. Set the rules aside. Read the revision aloud and ask: where does it sound like a person working through a problem, and where does it sound like a model completing a writing task? If a passage still feels templated after sentence-level repair, the issue is structural — do not force personality onto it. In rewrite mode, leave the awkward faithful sentence; in audit mode, report the structural finding.
- Consider no-op. If the source is already natural and fit for purpose, leave it alone. One weak signal is not enough to justify a rewrite.
Hard boundaries
- Do not invent facts, examples, citations, dates, measurements, product capabilities, feedback, emotions, opinions, or first-person experience.
- Do not turn vague attribution into direct fact, or uncertainty into certainty.
- Do not change the actor, owner, object, direction, completion state, causal relationship, comparison, or type of effect.
- Keep quotations, titles, proper names, technical terms, commands, paths, identifiers, logs, schemas, and literals exact unless the user explicitly asks to edit them.
- In file mode, preserve frontmatter, code blocks, inline code, structured data, and link destinations. Edit link labels only when prose editing requires it.
- Treat unsupported claims as source content, not permission to fact-check. Preserve their attribution and confidence. In audit mode, note that a source is missing; in rewrite mode, do not add one or silently promote the claim.
- Never fetch or infer private voice samples. Use only material the current user explicitly supplies for the current task, and do not persist it. A separate author preference layer is valid only when the user has explicitly maintained it and the current target is in its declared scope.
Positive target
Aim for prose that:
- says the useful thing without announcing that it is about to say it;
- uses the register the situation calls for, including neutral technical prose;
- varies rhythm when natural, without forced fragments or manufactured punch;
- preserves unusual, specific, defensible details already in the source;
- keeps uncertainty, mixed feelings, asides, repetition, and rough edges when they belong to the author;
- can be read aloud without sounding like a presentation script.
“More human” does not mean more casual, more emotional, more first-person, or less precise. It means less statistically generic and more faithful to the author, audience, and purpose.
Output contracts
Rewrite
Return one recommended revision. Do not show a draft, score, internal ledger, or self-audit. Add a brief note only when a material ambiguity or fidelity risk cannot be resolved from the source. In long or mixed-quality text, edit only the parts that need it and return the rest unchanged.
Audit
Return only material findings, normally one to five. For each finding include:
- the exact excerpt;
- the pattern and why it matters here;
- the fidelity or register risk, when one exists;
- a concrete editing direction.
Do not flag isolated words without contextual evidence.
If there are no material findings, say so and do not invent issues to fill the format.
File
Edit only the requested file. If the final prose is unchanged, do not rewrite the file. Report the path and a short summary instead of pasting the full file.
No-op
In rewrite or embedded mode, return the exact input without a preamble. In file mode, retain the file byte-for-byte and report that no change was needed. In audit mode, say there are no material findings.
Other versions108 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 108
- molt-bot/clawdbot★ 391ksame name
- every-app/open-seo★ 26ksame name
- millionco/react-doctor★ 15ksame name
- jaredpalmer/kev★ 9.8ksame name
- joe-bell/cva★ 6.9ksame name
- sanity-io/sanity★ 6.4ksame name
- millionco/expect★ 3.6ksame name
- rohitg00/pro-workflow★ 2.9ksame name
- compozy/compozy★ 2.8ksame name
- Jwuthri/Tracely-ai★ 1.5ksame name
98 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
Used alongsideProjects with it also use control-cli, verify-this and 10 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.
ElyzeSolutionsSkillcontrol-cli
Build or adapt a local harness to drive, inspect, and profile an interactive CLI or TUI without external services. Use for CLI UX checks, startup regressions, memory leaks, hangs, prompt flows, or terminal demos.
Also in 6 projects that use this
learnerLjSkillverify-this
Verify a claim with fresh local evidence: restate it falsifiably, capture baseline and treatment, compare artifacts, and return VERIFIED, NOT VERIFIED, or INCONCLUSIVE.
Also in 6 projects that use this
JwuthriSkillsimple-issue-description
Turn a rough bug report, feature request, support note, or pull request into a short, plain-language issue focused on the problem and desired behavior. Use when a contributor asks to simplify an issue, explain what a PR is for, create the corresponding issue for a PR, remove implementation detail from a report, or invokes /simple-issue-description.
Also in 3 projects that use this
JwuthriSkillopenseo-review-web-content
Write and review content for the OpenSEO website (web/) — blog posts, guides, feature pages, FAQs. Distills the philosophy for on-brand, useful, accurate content. Use whenever adding or editing user-facing prose in web/content or web/src.
Also in 3 projects that use this
ElyzeSolutionsSkillcontrol-ui
Build or adapt a local browser/CDP harness to drive and inspect a web, IDE, or Electron UI. Use for local UI verification, screenshots, accessibility snapshots, perf profiles, visual diffs, or reproducing UI bugs.
Also in 7 projects that use this
jaydubya818Skillpoteto-mode
Use when non-trivial work should follow the smallest rigorous Jstack workflow.
Also in 6 projects that use this
- +2 more from Jwuthri
In this repository28 other files it gives its agents.
- antigravity-coding-agentSkill
- attention-frontierSkill
- audio-transcriptionSkill
- brainstormSkill
- build-verifierSkill
- claude-code-coding-agentSkill
- codex-async-followupSkill
- codex-coding-agentSkill
- codex-session-controllerSkill
- codex-session-namingSkill
- cursor-coding-agentSkill
- debuggingSkill
- explainSkill
- explore-and-planSkill
- faster-learning-coachSkill
- frontend-harness-slidesSkill
- generate-commit-messageSkill
- git-task-workspaceSkill
- grok-coding-agentSkill
- home-config-syncSkill
- llm-wiki-captureSkill
- local-file-organizerSkill
- long-task-controlSkill
- parallel-goal-workflowsSkill
- reverse-engineerSkill
- review-and-improveSkill
- tddSkill
- x-twitter-readerSkill
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
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.

