humanize — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited humanize (Agent Skill) and scored it 87/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 3 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 3 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Transform stiff, AI-flavored, or over-polished writing into something that sounds like it came from an actual person -- with a distinct voice, natural rhythm, and the small imperfections that signal authenticity.
The user provides text to humanize. They may also provide:
If none of these are provided, infer them from the text. Ask only if the target voice is genuinely ambiguous and would change the output significantly.
If the user has placed writing samples in references/voice-profile/my-voice.md, read that file before rewriting. It contains their preferred sentence patterns, register, and vocabulary. Apply those patterns instead of projecting a generic voice.
Before rewriting, read the text as a human editor would. Identify internally:
detail, or a light edge that has been smoothed away
Do this quickly and internally. Do not narrate this to the user unless asked.
For a full catalog of AI patterns and their fixes, read: references/patterns.md
Apply these principles in order:
Voice first, fixes second. Do not just remove bad patterns. Replace them with something that has character. A human does not just avoid em dashes -- they use shorter sentences, or fragments, or a question.
One idea per sentence, usually. Most AI text over-compounds. Break it up. Short sentences are not unsophisticated -- they are confident.
Concrete over abstract. Replace vague abstractions with specifics wherever possible. "Improved performance" becomes "cut load time in half." "Valuable insights" -- name one. If there is no specific to reach for, cut the claim.
Preserve the actual content. Never add claims, details, or assertions not present in the source. Never invent examples. If something needs elaboration, flag it for the user -- do not fabricate it.
Register calibration. Match the register to the channel. Read references/channel-rules.md for channel-specific constraints (LinkedIn, email, blog, Slack). Character limits in that file are hard limits.
Natural emphasis. Real writers use bold and italics sparingly and only for genuine emphasis. If the text over-emphasizes, strip it.
Run these mentally before returning output:
Any phrases you would never hear in speech?
triads? Any hollow pivots (see references/patterns.md)?
Did any meaning get lost?
any character or length limits?
If any of these fail, fix before returning.
Default: Return the rewritten text only -- no preamble, no explanation, no "Here is the humanized version:". Just the text.
Add a note only when:
(Removed: "X" -- no supporting detail in source. Add if you have one.)
When the user asks for explanation: Walk through 2-3 of the most significant changes and why. Keep it brief. Focus on what changed in voice and rhythm, not just what was deleted.
"Make it sound like me" -- Read references/voice-profile/my-voice.md first. Extract voice patterns: sentence length, how they open paragraphs, whether they use fragments, their preferred register, recurring vocabulary. Apply those patterns. If no profile exists, ask the user for 1-2 examples of their writing before rewriting.
Minimal AI contamination -- If the text is mostly fine but has a few patterns, do targeted surgery rather than a full rewrite. Preserve what is working.
Heavy AI contamination -- If nearly every sentence has a pattern, do a full rewrite. Do not try to patch it. Rewrite from the core message outward.
Soulless but technically correct -- The fix is not removing patterns. It is adding a point of view, a specific detail, or a human beat. Make a small creative choice and flag it if needed.
Formal writing (legal, regulatory, academic) -- Some register is intentional. Do not over-humanize. Focus only on removing AI patterns that would not appear in competent human writing of that type.
Read these when needed (Progressive Disclosure -- do not load all at once):
references/patterns.md -- Full AI pattern catalog with bad to good rewritesreferences/channel-rules.md -- Channel-specific rules and hard limitsreferences/voice-profile/README.md -- How users set up their voice profilereferences/voice-profile/my-voice.md -- User's personal voice samples (if populated)~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.