Open source An Agent Skill
anti-slop-slop-canon removes the habits that make writing sound
machine-made, then learns a transparent, editable profile of your voice. One
profile for written and spoken content.
Every example below is real output from the project’s own rules.
What’s left your point, in your own voice.
Out of the box it strips hype, canned framing, and ornamental punctuation without inventing a claim the source never made.
Same facts. Removed the hype, a manufactured contrast, and an exclamation point.
No em dash, no semicolon, no exclamation point, no mid-sentence colon.
Without the skill, an assistant leans on hype and ornamental punctuation. The bundled defaults remove those habits from any draft. Add your profile and the same facts come out sounding like you.
What a typical assistant writes
We're thrilled to announce that v2.4 is here, and it's a game-changer! Here's the thing—this groundbreaking release makes your whole workflow feel seamless. The CLI now has a powerful new --watch flag, cold starts are blazing fast (down from 4.2s to a mere 1.8s), and we've finally squashed that pesky Windows bug that kept dropping your environment variables. You're going to love it!
Cleaned with the bundled defaults
Version 2.4 ships Thursday. The CLI gets a --watch flag. Cold start drops from 4.2 seconds to 1.8 seconds. The Windows bug that dropped environment variables is fixed.
In a sample personal voice
2.4 lands Thursday. The CLI finally gets a --watch flag. Cold start went from 4.2 seconds to 1.8. And that Windows bug that ate your environment variables is gone. Ship it.
The same profile covers what people read and what they hear.
Blog posts, docs, and social posts
Before
This isn’t a redesign. It’s a reset.
After
This redesign changes the navigation and search model.
A blog opener with the contrast framing removed, from the bundled defaults.
Video scripts, narration, voice-agent replies
Before
Deploying takes three steps:
After
Deploying takes three steps. First, pull the latest main. Then run the database migration. Last, restart the workers.
The list numbers drop out, so they are never read aloud. The order stays the same.
You give it a few samples. It finds the traits that are your voice, compiles a profile you approve, and writes in your voice from then on.
Detected from a writing sample
Kept as your voice
Set aside · subject matter
Set aside · one-off quirks
The same sentence, before and after
Default voice
The cache warm-up now finishes in about two seconds.
Your voice
Cache warm-up is down to about two seconds now. Good enough.
A short, adaptive interview. One question at a time.
Skip setup and use the bundled defaults right away.
Not now. Use the defaults and ask again later.
The compiled profile is a plain Markdown file in your project, not inside the installed skill. Open it and edit any rule. The skill leaves your edits alone.
Dry, direct, first person. Short sentences and clipped fragments. Concrete nouns over adjectives. Sounds like a builder talking to other builders.
Lead with what shipped. One idea per line. Fragments are fine when they land. Close on the point, not a summary.
Read it the way you would say it. Drop list numbers when the text is read aloud. Keep sentences short enough to follow on first listen.
Prefer: ship, land, fix, cut. Avoid: seamless, robust, leverage, delve.
Short sentences, shorter when the point is sharp. A longer line only when a detail earns it.
A one-line sign-off. Usually “Ship it.”
No “here’s the thing.” No em dashes. No three-part lists added for rhythm.
Ask for the realtime prompt and the skill returns a compact, style-only module for a voice agent’s system prompt. It shapes delivery only, so it sits beside your agent’s own instructions rather than replacing them.
Speak with calm confidence in plain, conversational language. Lead with the useful point. Keep the tone capable and restrained. Prefer concrete names, actions, and quantities to broad praise. Use familiar words and contractions. Let light, dry humor appear only when it fits the moment.
Return only words intended to be heard. Use no Markdown, emojis, bullets, tables, headings, visual notation, stage directions, or inaudible instructions. Convert symbols and abbreviations into natural spoken wording when their pronunciation could be unclear.
Write concise, pronounceable sentences that make sense on the first hearing. Put context next to the claim it explains. Break up dense clauses, but vary sentence length so the delivery stays natural. Repeat a key noun when a pronoun could confuse the listener. Use transitions only when they help the listener follow the thought.
Avoid hype, marketing jargon, canned openers, canned closers, stock transitions, filler intensifiers, and unsupported enthusiasm. Do not manufacture contrast formulas or three-part slogans. Start and end on substance.
Preserve warranted uncertainty. State limits directly. Read the response aloud in your head before returning it. Fix anything difficult to pronounce, easy to mishear, or dependent on seeing the text.
Installation uses npx skills. A project copy keeps its own state
and never falls back to the global one.
Activates the skill in every project on the machine.
npx skills add codeSTACKr/anti-slop-slop-canon -g Scoped to one repository, with its own isolated state.
npx skills add codeSTACKr/anti-slop-slop-canon It works with coding agents that read Agent Skills, including Claude Code, Cursor, and Codex.
This skill will not help you evade AI-detection tools and makes no claim that its output is undetectable. Its purpose is to remove habits that make writing sound generic and to help you sound like yourself.