AI Writing Patterns

AI Writing Patterns

Hold Your Voice keeps two layers separate:

• The 220-pattern catalog is the complete editorial reference, with examples, spotting guidance, ugly versions, and legitimate exceptions.
• The local CLI uses only a small, deterministic subset. Every executable rule has a stable ID, severity, repair direction, a positive test, and a counterexample test.

A match is an editorial signal, not proof that AI wrote a sentence. Human writers can use these patterns deliberately; models can avoid them. Read the line in context.

Current executable rules

Rule	Severity	Example	Repair direction
ai.delve	Red	“We will delve into the results.”	Name the action: inspect, compare, test, or explain.
ai.leverage	Red	“Leverage your data.”	Say what is being used and how.
ai.tapestry	Red	“The tapestry of modern work.”	Name the actual relationship.
ai.holistic	Red	“Take a holistic approach.”	Describe the included steps or scope.
ai.robust	Yellow	“A robust framework.”	State the evidence for strength.
ai.landscape	Yellow	“The marketing landscape.”	Name the field, actors, or change.
ai.game-changer	Red	“This is a game-changer.”	Explain exactly what changed.
ai.formulaic-connector	Yellow	“Firstly, review the file.”	Move directly to the next point.
ai.hedging	Yellow	“Perhaps this matters.”	Make, qualify, or remove the claim.
ai.signpost	Yellow	“This is why it matters.”	Lead with the observation.
ai.not-just	Red	“Not just fast, but reliable.”	State the two supported facts plainly.
ai.truth-setup	Yellow	“The hard truth is…”	Name the fact without a manufactured reveal.
ai.em-dash	Yellow	“The logs failed — retry later.”	Use a sentence break when it improves clarity.
ai.question-hook	Yellow	“Have you checked the logs?”	Start from a concrete observation.
ai.abstract-cluster	Yellow	“Alignment and clarity drive value.”	Use concrete nouns and actions.

Red findings block the AI Editor gate. Yellow findings ask for review and never trigger an automatic rewrite.

How to use the catalog

1. Start with the sentence, not the label. Ask what the wording does for the reader.
2. Look for repetition, clusters, low specificity, or a mismatch between the language and the situation.
3. Keep a pattern when it has a specific, necessary job. The full catalog documents legitimate exceptions and good uses.
4. Do not add a catalog item to the CLI just because it sounds familiar. It needs a narrow detector, a severity policy, and tests that show both the match and the counterexample.

Run node dist/cli.js patterns to inspect the exact rule set installed in a checkout.