Writing · Voice · Safety
Writing without the AI sheen.
The goal is not to disguise machine-written text. It is to help a person express their own observation, uncertainty, vocabulary, and point of view more clearly—then keep ownership of the final pass.
01
Use AI as an editor, not a ghostwriter
Start with the thing you actually think: an observation, a tension, an example, or an unfinished argument. Ask for a clearer structure, not a generic “engaging post.” A useful editor preserves the writer’s vocabulary and uncertainty rather than replacing them with polished, interchangeable prose.
An awkward but real phrase can be better than a smoother line that could have been written by anyone.
This matters especially for professional social writing. The reader should encounter a person’s actual judgment—not a manufactured lesson, grand claim, or engagement hook.
02
A practical editing loop
- Write the raw idea. Include the example, qualifier, and opinion that make it yours.
- Build a voice reference. Use your own past posts, emails, or notes—not a generic personality label.
- Draft with constraints. Ask the model to keep your claims, avoid invented details, and leave out a call to engagement unless you supplied one.
- Run a separate AI-ism audit. Identify patterns first; make only the changes you agree improve the piece.
- Read it aloud and own it. Restore phrases, rhythm, or uncertainty that sound like you. Do not publish a sentence you would not say.
This separation matters. A linter can help spot repetitive rhetorical moves; it should not decide what the writer means.
03
Patterns worth watching
Look for clusters and context, not a detector score. Many of these habits occur in ordinary human writing too.
- Inflated importance, generic praise, or a conclusion that simply restates the point.
- Repeated “not X, but Y” turns, tidy rules of three, or identical sentence and paragraph lengths.
- Generic hooks, artificial suspense, fake quotes, or a rhetorical question added only to invite comments.
- Corporate filler, unexplained jargon, and adjectives that announce significance instead of supplying evidence.
- Artificial imperfection—typos, slang, or fragments inserted solely to fool a classifier.
Avoid optimizing for “passing” an AI detector. No such checklist establishes authorship, and detector-driven edits can erase a writer’s real style.
04
Four skills, reviewed before installation
NVIDIA SkillSpector v2.11.2 ran in static-only mode against the exact commits below. We also ran OWASP’s high-signal direct prompt-injection patterns; none of the four targets matched those narrow expressions. That is triage evidence, not a safety certificate.
| Project | Static result | Review status |
|---|---|---|
| Avoid AI Writing | 99 · Critical · Do not install | Not installed |
| Humanizer by blader | 47 · Medium · Caution | Not installed |
| Humanizer Skill by Aboudjem | 45 · Medium · Caution (targeted skill scan) | Not installed |
| Ghostwriter | 39 · Medium · Caution | Not installed |
Avoid AI Writing: static evidence and human read
It offers detect, rewrite, and in-place edit modes, and explicitly says that AI-writing patterns are not authorship proof.
Exact commit: 64b2de64239e09d87267234edb4589ef0b5aff5d
Scan: 99 · Critical · Do not install · 58.3% static coverage · partial
What needs attention: The score is dominated by incomplete reference analysis and contextual matches: visible HTML boundary comments, the phrase “output rules” in documentation, and Unicode normalization used by the detector. The missing tool declaration is real, and bundled Node.js helpers still need code review. Edit mode can change a file, so use detect mode first and confirm an exact file scope before any write.
Humanizer by blader: static evidence and human read
A prose-rewriting skill based on Wikipedia’s Signs of AI writing that asks the agent to retain claims, avoid invented facts, and follow a supplied sample’s voice.
Exact commit: 9862685f575c65a8247f90369951df1b3416e3d6
Scan: 47 · Medium · Caution · 100% static coverage
What needs attention: The scanner flagged no declared tool scope, unpinned npx examples in maintainer documentation, and the phrase “without warning” inside a before-and-after example. The examples are not evidence of an active jailbreak, but the project still needs a semantic review before adoption.
Humanizer Skill by Aboudjem: static evidence and human read
The focused scan covered the installable humanizer folder rather than the repository’s documentation and site assets. It declares read, write, edit, search, and user-question tools.
Exact commit: a58df065367550b6ce40ff3f648335018d8e0589
Scan: 45 · Medium · Caution (targeted skill scan) · 83.3% static coverage · partial
What needs attention: The scanner treated “without warning” and “do not judge” in prose examples as anti-refusal language, and misread a documented copy command as snooping and persistence. Those are contextual false positives. More importantly, the skill automatically loads a project-level humanizer-context.md file as voice guidance. Treat that file as untrusted unless the user explicitly selected it; it can otherwise become an indirect instruction source.
Ghostwriter: static evidence and human read
A four-skill voice-profile collection for analyzing samples, creating or blending profiles, and applying a profile to drafts.
Exact commit: 6245570aedebb0a10c6e801c6d4c9ee7fa29484e
Scan: 39 · Medium · Caution · 100% static coverage
What needs attention: The scanner flagged the real “no disclaimers” instruction in the apply skill; two other persistence findings were false positives on the ordinary word “create.” Profile loading across project and user directories still broadens what can influence a rewrite, while voice samples may be sensitive. Limit reads to a consented profile and never use a writing-style rule to suppress safety-relevant disclosures.
A static scan does not verify a publisher, execute every path, establish runtime behavior, or remove false positives and false negatives. The Humanizer Skill’s repository-wide preliminary scan was critical because of documentation and bundled assets; its focused installable folder scanned low. Scope changes the result, so adoption must be tied to an exact artifact.
05
skills.sh registry check: one warning, the rest pass
Retrieved September 15, 2026. Exact skills.sh listings exist for Avoid AI Writing and both Humanizer projects. Ghostwriter is published as four separate skills, so each listing was checked instead of inventing a repository-wide verdict. Every row links to the listing; each provider result links to its detailed audit.
| Published skill | Gen Agent Trust Hub | Socket | Snyk | Audited |
|---|---|---|---|---|
| Avoid AI Writing | Pass · Safe | Pass | Pass · Low | Sep 15, 2026 |
| Humanizer by blader | Pass · Safe | Pass | Pass · Low | Sep 7, 2026 |
| Humanizer by Aboudjem | Pass · Safe | Pass | Pass · Low | Sep 14, 2026 |
| Ghostwriter · voice-apply | Warn · Medium | Pass | Pass · Low | Jun 26, 2026 |
| Ghostwriter · voice-analyze | Pass · Safe | Pass | Pass · Low | Jun 26, 2026 |
| Ghostwriter · voice-create | Pass · Safe | Pass | Pass · Low | Jun 26, 2026 |
| Ghostwriter · voice-blend | Pass · Safe | Pass | Pass · Low | Jun 26, 2026 |
Avoid AI Writing: three passes, with declared exposure
Gen Agent Trust Hub records bundled local Node.js scripts and indirect-prompt-injection exposure, but also notes explicit instructions to treat audited prose as data. Socket reports no flagged malicious, trust, obfuscation, or autonomy behavior; Snyk reports Low risk with no issues detected.
Snapshot note:Socket's package identifier ends in 64b2de64239e09d87267234edb4589ef0b5aff5d, which matched the exact commit in the static-review table above.
The two Humanizers: three passes do not erase the input boundary
Gen Agent Trust Hub marks both Safe while documenting the same material weakness: each skill processes untrusted prose without strict boundary markers or sanitization. Aboudjem's version also declares Read, Write, Edit, Grep, Glob, and user-question capabilities, so embedded instructions could influence a higher-impact agent if the host does not enforce separation.
Snapshot note:blader's Socket identifier matched the exact reviewed commit 9862685f575c65a8247f90369951df1b3416e3d6. Aboudjem's identifier is a content hash, not a verified Git commit mapping; the exact snapshot correspondence remains unproven.
Ghostwriter: provider disagreement changes the decision
Gen Agent Trust Hub rates voice-apply Warn / Medium because it suppresses AI-identity and limitation disclaimers and accepts untrusted text without explicit boundaries. It also notes an indirect-injection surface in voice-blend, which loads project and user voice profiles and can run a local Python helper. Socket passes all four listings, while Snyk rates each Low with no issues detected.
Snapshot note: the four registry entries use content hashes that could not be mapped to Git commit6245570aedebb0a10c6e801c6d4c9ee7fa29484e. Treat the June audits as useful but not exact-artifact evidence.
The Medium warning blocks a blanket approval of Ghostwriter, and the Humanizer passes still expose indirect prompt-injection surfaces. These results support the local findings above; they do not replace them or turn any project into a verified-safe install.
This skills.sh registry check reflects third-party audits of mutable registry snapshots. It does not verify publisher identity, the exact local artifact, dependencies resolved at install time, or runtime behavior. A missing or unmapped snapshot remains unknown.
06
Test safely before you trust it
Keep a first trial narrow: use non-sensitive text, block network access unless the workflow needs it, and start in detect-only or copy-out mode. Do not give a writing skill access to a project, profile directory, or sample corpus until you know exactly what it reads and writes.
The static, registry, and local semantic reviews all identify areas that need caution. Any later approval should pin the exact commit, define allowed tools and files, preserve a rollback copy, and require a fresh review for updates.
07
A reusable editor brief
Use this as a starting point, then replace the generic terms with your own preferences and examples.
Act as my editor, not my ghostwriter. Preserve my ideas, opinions, anecdotes, uncertainty, vocabulary, and point of view. Improve clarity and structure without adding a lesson, engagement question, or made-up detail. Prefer specific observations over generalized insights. Flag familiar AI writing patterns, but make minimal edits and leave passages that already sound like me alone.