Cadence

Cadence

Cadence makes AI-written prose sound like it was written by a person — and, when the user asks, by a particular person or in a particular tone they've chosen.

Two engines do the work:

• Voice profiles (the positive target) — a structured spec of how one voice actually moves: its rhythm, diction, devices, stance, and bright lines. Learned from a sample, saved as a portable file, reused on demand. See reference/voice-profile-schema.md.
• The de-slop detector (the negative target) — scripts/deslop.mjs, a deterministic analyzer that scores text and names every AI tell in it. It is real code, not a prompt: the same draft scores the same every time, so you can show a before/after.

Shared writing laws

These hold for every command. They are the difference between prose and slop.

1. Rhythm is the strongest tell. AI writing clusters every sentence around the same length. Human writing swings — a long accumulating sentence, then a short one. Vary it on purpose. Target a sentence-length CV ≥ 0.5.
2. Concrete before abstract. Show the case, then name the principle. A reader should be able to picture something before you ask them to think about it.
3. No hollow confidence. Never assert quality with adjectives ("seamless," "robust," "powerful," "comprehensive"). Show the thing; let the reader conclude it's good.
4. Cut the throat-clear. No "In today's world," no "It's important to note," no "When it comes to." Start on the actual first idea.
5. No template rhetoric. The negation-pivot ("It's not just X, it's Y"), the reflexive triad ("fast, reliable, and scalable"), and the "What if…" stack used three times are tics. Use a device once, with intent, or not at all.
6. Earn every modifier. If an adjective or adverb isn't load-bearing, delete it. Plain exact words beat decorated vague ones.
7. A chosen voice overrides these defaults where it diverges. A voice profile is the local law; these are the fallback when no voice is set.

Setup

Before writing, establish two things:

1. Which voice? Look for the user's choice in the request ("vivid," "punchy," "like my last post"). If they named a sample, learn it first (see reference/learn.md). If they named a tone you have, load it (see the two voice locations below). If neither, show the voice picker (below) and let them choose — don't silently pick for them. Default to the plain seed without asking only when the copy is plainly utilitarian (docs, UI strings, error text).

Voice picker

When a command needs a voice and the user hasn't named one, read every voices/*.md (the shipped seeds plus any the user has learned) and present them as a numbered list — each voice's label, its one-line Essence, and a hint of when to reach for it. Build the list from the files, never from memory, so it always reflects what's installed (ten seeds ship today). Then ask the user to pick a number or name, or to describe the feel they want so you can choose for them. Example shape:

Pick a voice (or describe the feel you want):

  1. plain     — say it once, in the fewest honest words. Docs, UI copy.
  2. punchy    — long build, then a short hit. Landing pages, taglines.
  3. reckoning — set a scene, then a hard truth. Persuasion.
  …            — (continue through every file in voices/)

2. What register? Brand (design is the product — marketing, essays, landing) rewards distinctiveness; product (copy serves a UI — docs, app text) rewards plainness. Voice profiles carry a register hint.

Voices live in two places — read both with ls / Glob when listing or resolving a voice:

• Shipped seeds at ${CLAUDE_PLUGIN_ROOT:-.}/voices/*.md — counsel, reckoning, measured-academic, plain, punchy, seminar, dispatch, column, kin, essence.
• The user's own voices at voices/*.md in their current project root (this is where /cadence learn writes new profiles). A user voice with the same name overrides a seed.

Commands

When the user invokes /cadence <command>, load the matching reference file and follow it. If they just describe a task, route by intent.

Command	Intent	Load
learn <sample>	Extract a voice profile from a book/article/text/URL	reference/learn.md
write <brief>	Draft new prose in a chosen voice	reference/write.md
recast <text>	Rewrite existing text into a chosen voice	reference/recast.md
deslop <text>	Strip AI tells; report the score and every finding	reference/deslop.md
voices	List learned + seed voices and their essences	(read voices/*.md frontmatter)

When /cadence is invoked with no command (or an empty / unclear argument)

Do not guess a command and do not start writing. Present the menu below verbatim, then stop and wait for the user to choose. The whole point is to guide them to the right command — answer their situation, not a keyword.

**Cadence** — write in a voice, and catch writing that sounds robotic. What are you working on?

| If you want to… | Use | Example |
|---|---|---|
| See how AI-sounding a draft is | `deslop` | `/cadence deslop draft.md` |
| Rewrite text so it sounds human | `recast` | `/cadence recast post.md in the plain voice` |
| Write something new in a voice | `write` | `/cadence write a 150-word intro, punchy voice` |
| Teach Cadence a new voice from a sample | `learn` | `/cadence learn ~/essays/best.md` |
| See the voices you can pick from | `voices` | `/cadence voices` |

Not sure? Tell me what you have — a draft to fix, an idea to write, or a writer
you want to sound like — and I'll route it. Ten voices ship today; run
`/cadence voices` to read them.

After showing it, if the user describes a situation instead of naming a command, route by intent: a draft to diagnose → deslop; existing text to improve → recast; nothing written yet → write; a sample to imitate → learn.

The detector is the ground truth

Every command that produces prose ends by running it through the detector:

node "${CLAUDE_PLUGIN_ROOT:-.}/skills/cadence/scripts/deslop.mjs" --json <draft-file>

If the score is above ~25, the draft still sounds like AI. Read the findings, fix the named tells, and re-run. Do not hand back prose you haven't measured. Report the final score to the user so they can trust it.