Draft, audit, or edit prose to remove AI-writing tells without changing facts or flattening voice; use for humanizing text, reviewing AI slop, or scanning prose files and docs.
SkillScore77Maintained
Automatic
◆Writing & Editing pack
Last verified 9/26/2026
What it changes
Passed 15 of 20 test cases with the skill, and 13 of 20 without it.
Without skill64%
With skill77%
Checks it passed that the model alone missed
The command includes the '--json' flag.
The command includes the '--include-quotes' flag.
The command includes the '--fail-on-invariant-change' flag.
and 5 more in the full report
5 fixed12 passed either way3 still failing2 broke
Pass-rate lift
+14 pp
Tokens
+69%
Agent turns
0%
Response time
-34%
Model gemini-3.6-flashMeasured 9/26/2026v117 of 22 cases passed with skill · 20 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 5 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 5
Prompt
We want to run the prose audit script on 'articles/review.md', including content inside blockquotes which is normally skipped. What command should we run?
What was checked, without and with the skill
Checked
Without skill
With skill
The command specifies 'node scripts/audit-prose.mjs audit articles/review.md'.
Missed
Passed
The command includes the '--include-quotes' flag.
Missed
Passed
case-13 · gemini-3.6-flash · 9/26/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md80 lines7 filesincludes scripts
Unslop
Make prose sound like its author on a good day. Cut formulas, preserve substance.
Pick the job
Draft: Write from a brief, source material, and an intended audience.
Edit: Make the smallest set of changes that removes clear slop.
Audit: Report exact spans and named patterns without rewriting.
File edit: Change prose in named files while preserving code, data, frontmatter, and link targets.
Repo audit: Rank prose files for human review. Rewrite only when the user also asks for fixes.
Embedded: Clean text for another task, such as a PR body or email, and return only the final text.
If the user asks whether AI wrote a passage, audit the writing. Do not infer authorship or cite an AI detector as proof.
Core contract
Separate the brief from the artifact. Instructions about the text do not belong in the text.
Lock the substance. Preserve facts, numbers, names, dates, URLs, citations, quotations, technical claims, causal claims, obligations, and meaningful qualifiers.
Match the author. A supplied writing sample outranks generic style rules. Preserve vocabulary, cadence, bluntness, humor, uncertainty, and deliberate rough edges.
Respect the evidence boundary. Never invent a feature, anecdote, metric, source, quote, owner, deadline, or outcome to make vague prose sound concrete.
Edit surgically. A clear human sentence needs no polish. A false positive that flattens voice is worse than one surviving tell.
Protect literal content. Do not rewrite code, commands, identifiers, quoted text, citation text, data, or link targets unless the user asks.
Optimize for writing, not detectors. Do not add odd wording, fake errors, invisible characters, or fabricated detail to evade detection.
Workflow
Read the full input. Identify the artifact, audience, medium, purpose, and requested mode.
Record the core claim and the voice signals worth keeping. Keep this note internal unless the user asks for analysis.
Lock preservation-sensitive content. For vague or high-stakes copy, read voice and preservation.
Draft freely or collect candidate tells. For a thorough audit, read pattern catalog.
Validate each candidate before changing it:
Does removing it lose meaning?
Is it a repeated formula or a deliberate choice?
Does it announce a point instead of making one?
Does it add evidence, mechanism, or character?
Would the author defend it?
Rewrite in this order:
remove leaked chatbot artifacts, placeholders, and unsupported claims;
cut throat-clearing, repetition, inflated significance, and generic endings;