legendaryslop

A portable agent skill that rewrites English and Russian prose to read like a named person wrote it, without changing what the text says.

Most "humanizer" tools swap words against a blocklist and stop there. That trades one fingerprint for another: no flagged phrases, but also no voice, paragraphs cut to the same length, every list holding exactly three items. legendaryslop treats each pattern as a weighted signal, acts on clusters instead of single words, and stops before it flattens the writer.

The problem it solves

AI-drafted prose fails in two directions at once. It carries tells: throat-clearing openers, "it is important to note", rule-of-three padding, a "Challenges" section that resolves into optimism. Strip those with a find-and-replace pass and you get the second failure: even, voiceless text that a reader still clocks as machine-made.

The fix is editorial judgment, not a word list. legendaryslop encodes that judgment as a workflow: lock the facts, find where several tells cluster, rewrite those spots for a real reader, and leave working sentences alone.

Before and after

Illustrative, not a benchmark:

Before. In today's fast-paced digital landscape, it is important to note that clear communication plays a pivotal role in fostering team alignment. Clear, concise, and consistent messaging is key.

After. Teams drift when messages are vague. Say what changed, who owns it, and by when.

What it does

• Two languages, native rules for each. English and Russian guidance are separate. An em dash is a cluster-only tell in English and normative punctuation in Russian; the skill treats them differently instead of deleting both.
• Facts are locked before any edit. Claims, numbers, dates, quotations, citations, and modal distinctions (may, likely, must not, должен, вправе, нельзя) survive the rewrite. The skill never invents an anecdote, a statistic, or a source to sound human.
• Clusters, not tokens. One dash, one passive sentence, one "however" is not evidence. It rewrites where signals combine or a phrase carries no meaning.
• Depth matches the defect. A surface pass for surface problems; a structural rebuild only when local edits would leave the same template underneath.
• Genre-aware. Legal text keeps its modality and cross-references. Fiction keeps its fragments and unresolved tension. A postmortem keeps its verdict.
• Optional linter. scripts/lint_patterns.py flags candidate EN/RU patterns with confidence levels and per-language density checks. Every hit is a review candidate, never an automatic edit.

What it will not do

• Decide whether a human or a model wrote a text. It is not a detector.
• Promise to defeat an AI detector or prove authorship.
• Remove disclosure, provenance, or required attribution.
• Touch code, data, or verbatim quotations.

Install

The skill is one directory: SKILL.md plus references/, scripts/, LICENSE, and ATTRIBUTION.md. Any agent that loads Agent Skills can read it.

Claude Code

git clone https://github.com/FrankyFro/legendaryslop ~/.claude/skills/legendaryslop

Call it with /legendaryslop, or let the model reach for it on a "humanize" or "de-slop" request.

Codex CLI

git clone https://github.com/FrankyFro/legendaryslop ~/.codex/skills/legendaryslop

Any other agent

Point your skill loader at the cloned directory. SKILL.md is the entry point; the references/ files load only when a task needs them.

Usage

• Rewrite — paste text, ask to de-slop or humanize it, get the edited text back.
• Audit — ask for findings instead of a rewrite. Returns a table: confidence, excerpt, issue, smallest fix.
• File mode — point at a Markdown file. Prose is edited in place; code, frontmatter, tables, and link targets stay intact.
• Draft — no source text. The skill writes new prose from your brief through the same fact-first workflow.
• Intensity — light, balanced (default), or aggressive.

Linter:

python3 scripts/lint_patterns.py article.md --format json
python3 scripts/lint_patterns.py article.md --language ru

It classifies a file as one language; for mixed text, run it once per language.

How it works

SKILL.md sets a priority order that resolves conflicts: the user's intent first, then factual fidelity, then protected content, then a supplied writing sample, then genre and house style, and only last the skill's own pattern rules.

Guidance is split so the agent loads what a given task needs:

Credits

legendaryslop is a derivative work. It combines and adapts Humanizer by Siqi Chen, Stop Slop by Hardik Pandya, Humslop, and Sepia by Nanako Tsai, all under the MIT License. Full detail and the list of changes are in ATTRIBUTION.md. The original authors do not endorse this version.

License

MIT. See LICENSE.