A portable agent skill that rewrites or reviews text against one person's house style: a banned-phrase list for AI-sounding tells, plus a positive prose-structure guide for flowing, conversational technical writing. It is plain Markdown, so it can run in any harness that supports skill-style instructions.
Install with the cross-agent skills CLI:
npx skills add DeweyMarco/declankifyUpdate an existing install:
npx skills update declankifyTo install into every supported agent harness:
npx skills add DeweyMarco/declankify --agent '*'To target one configured harness, pass its agent name:
npx skills add DeweyMarco/declankify --agent <agent-name>Claude Code users can also install it as a plugin:
/plugin marketplace add DeweyMarco/declankify
/plugin install declankify@declankify
The skill is then invoked as /declankify:declankify.
Any agent harness can use the skill directly because the runtime artifact is
SKILL.md. Install it wherever your harness expects skill directories, or
copy SKILL.md into an existing skill folder.
For example:
git clone https://github.com/DeweyMarco/declankify.git /path/to/your/skills/declankifyOr, if you already have this repo cloned:
mkdir -p /path/to/your/skills/declankify
cp SKILL.md /path/to/your/skills/declankify/Invoke the skill however your agent harness exposes installed skills. Common forms include a slash command or a direct request:
/declankify
[paste your text here]
Please rewrite this in my style: [your text]
Two halves, one skill:
- Banned phrases and patterns — specific AI-speak tells to cut on sight: stock phrases ("here's the thing nobody tells you"), hedge-and-pivot constructions ("not just X, it's Y"), em dashes, forced rule-of-three lists, false ranges, and more.
- Prose structure rules — eight rules covering answer length, paragraph density (every sentence carries claim + mechanism + consequence), when to use headings vs. numbered lists vs. plain bullets, and how to open and close without formulaic scaffolding.
Most AI-detection guides only do the first half. This one also encodes what to do, not just what to avoid, so the output reads like a specific person rather than merely "not obviously AI."
Stock AI phrases and hedges: "here's the thing nobody tells you," "belt and suspenders," "you're absolutely right," "testament to," "landscape," "serves as" / "features" in place of "is" / "has," "let's dive in," "I hope this helps!," "in order to," vague unattributed sourcing ("experts believe"), and more — see SKILL.md for the full list.
| Pattern | Why it's banned |
|---|---|
| "Not X but Y" / "it isn't just X, it's Y" | Hedges the point instead of stating it |
| Colon-fronted assertion setup | Announces a claim instead of making it |
| Em dashes | Overused as a substitute for real sentence structure |
| Forced rule-of-three | Manufactures a triplet instead of using the natural count |
| Synonym cycling | Rotates words for variety instead of repeating the clearest term |
| False ranges ("from X to Y") | Substitutes a vague span for the actual list |
| Hyphenated buzzword pairs | "cross-functional," "data-driven" read as filler |
| Title Case Headings | Reads as marketing copy, not technical writing |
| Superficial "-ing" clauses | "symbolizing...," "showcasing..." manufacture false significance |
| Promotional adjectives | "breathtaking," "game-changing" |
| Stacked hedges | "might potentially perhaps" instead of one qualifier |
| Vague forecasts | "only time will tell" instead of a specific claim |
- Answer exactly what was asked, at the length it deserves — err short.
- Every paragraph and bullet carries claim, mechanism, and consequence together.
- Match the form to the content, and vary it: headings for comparison axes, numbered lists for sequences, plain bullets for parallel facts, paragraphs for reasoning.
- Don't shred connected reasoning ("because"/"so"/"but") into bullets.
- Open with the verdict and its central caveat, in plain sentences, not a bolded headline.
- Conversational but not dramatic: contractions, no scaffolding phrases, no theatrical labels.
- No compression: shortness comes from cutting content, not clipping sentences.
- End with a bottom line only when the answer weighed a real decision.
Before (AI-sounding):
Here's the thing nobody tells you about switching to a merge-on-read table format: it isn't just a performance optimization, it's a fundamental shift in how your data landscape operates. The solution serves as a testament to modern data engineering, offering breathtaking flexibility for teams navigating from ingestion to analytics.
After (Declankify):
Merge-on-read is cheap to write, but every read has to reconcile delete files against data files, so scans get slower and flakier until something compacts them, and now that's your problem to operate. If your workload is write-heavy and reads are rare, that trade is fine. If reads are constant, you're better off eating the write cost of copy-on-write instead.
- blader/humanizer — the AI-detection pattern list this skill's banned-phrase section builds on
- chitalian/offensive-ai-speak — a second independent blocklist of AI-speak phrases and structural patterns
- Wikipedia: Signs of AI writing — primary source behind both of the above
- 1.0.0 — Initial release: banned-phrase list plus the eight-rule prose structure guide, packaged as a portable skill and Claude Code plugin.
- 1.0.1 — Renamed from dewey-style to declankify. Added offensive-ai-speak to references.
MIT