CLClean Language

AI writing patterns

The habits that make AI prose painful to read.

Generated text often looks polished while saying less than it should. These patterns appear across Claude, ChatGPT, Gemini, Copilot, local models, corporate templates, and human writing influenced by them.

This is the actual guidance the skill uses. This page is the reader-friendly guide; the skill's full pattern catalogue is anti-patterns.md. Read the exact source on GitHub.

Punctuation used as theatre

Em dash overuseAI often inserts dramatic pauses where a comma, colon, parentheses, or separate sentence would read more naturally.
En dash misuseGenerated prose sometimes uses en dashes as decoration, substitutes them for hyphens, or places them where ordinary punctuation is clearer. Clean Language converts both marks to the punctuation the sentence needs, keeping them only inside quoted or verbatim material; ranges take a hyphen or “to”.
Colon followed by a sloganA normal point is inflated into a reveal, a punch line, or a 1-sentence conclusion.
Fragment stackingShort fragments appear in sequence to manufacture importance. Clear content does not need theatrical spacing.
Conjunction openersSentences begin with And, But, So, or Because as a rhythm device rather than a real connection to the previous sentence.

Formulaic contrast framing

One of the most recognizable AI habits is the repeated setup and reversal:

  • “It is not X. It is Y.”
  • “Not because X, but because Y.”
  • “The problem is not X. The problem is Y.”
  • “This is not merely X. It is a broader Y.”
  • “It does not just do X. It also does Y.”
  • “The question is not whether X, but how Y.”

Contrast is useful when the distinction matters. Repeated contrast framing creates artificial drama and delays the actual conclusion.

Common AI filler phrases

Throat clearing“It is important to note,” “it is worth mentioning,” “at its core,” “when it comes to,” “the reality is,” and “let us explore.”
Empty importance“This is critical,” “the implications are significant,” “the stakes are high,” “this cannot be overstated,” and “this is a game changer.”
Generic transitions“Moreover,” “furthermore,” “in addition,” “ultimately,” “on the other hand,” and “with that in mind” used mechanically between every point.
Automatic scene setting“In today’s fast paced world,” “in an ever evolving landscape,” “in the digital age,” and “as organizations navigate increasing complexity.”

Vocabulary fingerprints

No single word proves that a model wrote a passage. Clusters of predictable vocabulary can still make prose sound generic. Clean Language challenges words such as:

  • delve, tapestry, realm, landscape, journey, testament;
  • leverage, unlock, elevate, empower, foster, navigate;
  • robust, holistic, seamless, pivotal, crucial, transformative;
  • meticulous, nuanced, multifaceted, dynamic, comprehensive;
  • underscore, illuminate, resonate, reimagine, revolutionize.

These words remain valid when they carry specific meaning. The problem is automatic selection, repetition, and use as a substitute for evidence. A flagged word is a diagnostic signal, not proof of AI authorship. For example, “leverage” is empty jargon in “leverage synergies”. It is precise in “the fund reduced its leverage”, where it names a financial measure.

Structure that looks generated

3-part everythingEvery point becomes a trio of adjectives, nouns, benefits, risks, or imperatives because 3 sounds rhetorically complete.
Excessive headingsShort answers are split into many sections, each with a heading that repeats the content below it.
Bullet staircaseEvery sentence becomes a bullet, producing vertical length without adding structure or decision value.
Repeated conclusionThe introduction previews the answer, the body repeats it, the summary repeats it again, and the final line turns it into a slogan.
Question then immediate answer“Why does this matter? Because...” imitates a speech and often replaces a direct statement.
Uniform paragraph rhythmParagraphs have matching length, matching cadence, and identical concluding emphasis.

Tone people dislike

Fake enthusiasm“Great question,” “absolutely,” “you are exactly right,” and automatic praise that contributes no information.
Unrequested reassuranceEmotional cushioning is added to technical, operational, or factual answers where the reader needs accuracy.
Inspirational framingRoutine work becomes a journey, opportunity, transformation, or invitation to embrace change.
Artificial balanceEvery criticism receives a ceremonial positive counterpoint, even when it does not alter the finding.
Reader hand-holdingThe prose explains obvious implications, announces each step, and repeatedly tells the reader why a plain statement matters.

Meaning failures

Vague claimsThe text calls something significant, strategic, complex, or concerning without naming the evidence, consequence, owner, or decision.
Hidden ownership“The issue was addressed,” “the decision emerged,” and “mistakes were made” conceal who acted.
False certaintyAssumptions, estimates, and model guesses are written as verified facts.
Plain language damageTechnical, legal, policy, or standards language is simplified until the original requirement changes.
Unnecessary verbosityA simple answer becomes a long explanation because the model optimizes for completeness rather than usefulness.

Before and after

Each pair preserves every fact. Only the style changes.

Formulaic contrast

Before

This is not merely a technical fix. It is a critical improvement that will help ensure a robust and reliable service.

After

This technical fix will help make the service robust and reliable.

Repeated conclusion

Before

The migration is scheduled for Friday. The team will migrate the service on Friday. In summary, Friday is the planned migration date.

After

The service migration is scheduled for Friday.

Hidden ownership

Restore the actor only when the surrounding text names who acted. Where the actor is unknown, flag the gap rather than inventing one.

Before

The infrastructure team joined the incident call. The configuration was changed by the team, and this change resolved the issue.

After

The infrastructure team joined the incident call, changed the configuration, and resolved the issue.

What Clean Language does differently

Clean Language treats these patterns as diagnostic signals rather than universal bans, except the dash rule, which is absolute apart from its carve-outs. Each list names common instances, not the complete set; unlisted wording with the same function gets the same treatment. The standard preserves precise adverbs, useful passive voice, legitimate technical agency, required normative terms, accurate lists, and punctuation that improves comprehension.

Normative terms carry defined force, so Clean Language keeps them exactly. In policy and standards, “must”, “shall”, “should”, and “may” each set a different level of obligation, and changing one changes the requirement.

The governing rule is semantic preservation. Style cleanup cannot override facts, legal effect, technical accuracy, quoted language, contractual wording, policy obligations, or standards terminology.

Install Clean Language