natural-writing

Rylai Codex-Hermes-Claude Edition | Maintained and adapted by Rylai

Runtime Compatibility

• Codex: install under ~/.agents/skills/natural-writing.
• Hermes: install under ~/.hermes/skills/natural-writing or expose the bundle through skills.external_dirs.
• Claude Code: install under ~/.claude/skills/natural-writing or <project>/.claude/skills/natural-writing.
• Claude.ai and Cowork: upload and enable the matching per-skill ZIP.
• Resolve bundled files relative to this skill directory; do not depend on paths from another machine.
• Check tools, packages, credentials, network access, and runtime capabilities before execution.

Natural Writing

LLM text is detectable because it regresses to the mean: it replaces specific, surprising, concrete details with generic, positive-sounding language that could apply to almost anything. The subject becomes simultaneously less specific and more exaggerated — like a portrait fading from a sharp photograph into a blurry, generic sketch while someone shouts louder and louder that the person in it is important.

Every pattern below is a concrete manifestation of that tendency. The fix is almost always the same: be specific, be plain, and trust the reader.

---

1. Cut the Significance Inflation

The single most recognizable AI pattern is stuffing sentences with claims about how important, pivotal, or transformative something is.

What it looks like:

• "marking a pivotal moment in the evolution of..."
• "representing a significant shift toward..."
• "part of a broader movement across..."
• "highlighting the enduring legacy of..."
• "reflecting the transformative power of..."
• "contributing to the rich tapestry of..."
• "underscoring its role as a dynamic hub of..."

Why it happens: LLMs are trained on text where notable things are described with notable-sounding language. The model pattern-matches "thing being discussed" → "must emphasize its importance" regardless of whether the thing actually warrants it.

The fix: Delete the significance claim. State the fact. If the fact is significant, the reader will notice. If you have to tell them it's significant, it probably isn't — or you haven't presented the fact sharply enough.

Bad: "The library was established in 1962, marking a pivotal moment in the region's educational development and reflecting a broader commitment to knowledge accessibility."

Good: "The library opened in 1962."

If context is needed, provide it with a concrete detail, not an abstraction: "The library opened in 1962 — the first public lending library within forty miles."

---

2. Kill the Trailing Analysis

AI text habitually tacks on a participial phrase or subordinate clause at the end of sentences that "analyzes" what was just said, usually in vague terms about significance, impact, or recognition.

What it looks like:

• "...creating a lively community within its borders."
• "...further enhancing its significance as a dynamic hub."
• "...showcasing the brand's dedication to craftsmanship."
• "...demonstrating the ongoing relevance of his research."
• "...reflecting the influence of French rotary designs on German manufacturers."

The fix: End the sentence at the fact. If the analysis adds nothing a thoughtful reader couldn't infer, cut it entirely. If it contains a genuine insight, promote it to its own sentence and make it concrete.

Bad: "The dam generates 2,400 MW annually, underscoring the region's commitment to renewable energy infrastructure."

Good: "The dam generates 2,400 MW annually." Or, if the point matters: "The dam generates 2,400 MW annually — enough to power roughly 1.8 million homes."

---

3. Avoid the AI Vocabulary

Certain words spike in frequency in LLM output relative to human writing. One or two is coincidence. A cluster is a tell. Avoid overusing:

High-frequency AI words: delve, tapestry, multifaceted, nuanced, landscape (metaphorical), underscores, realm, foster, leverage (verb), pivotal, comprehensive, intricate, commendable, noteworthy, invaluable, meticulous, innovative, groundbreaking, cutting-edge, revolutionary, game-changer, holistic, synergy, robust, seamless, dynamic, vibrant, bustling, nestled, renowned, esteemed, testament

The principle: Prefer the shorter, plainer, more common word. "Shows" over "underscores." "Detailed" over "meticulous." "Useful" over "invaluable." "Complex" over "multifaceted." Often the fancy word can just be deleted — the sentence is stronger without it.

---

4. Use "Is" and "Are"

LLM text systematically avoids simple copulas (is, are, was, were) and substitutes longer constructions. Research shows a >10% decline in "is"/"are" usage in AI-era text. This is one of the subtlest but most measurable tells.

AI pattern → Human equivalent:

• "serves as a" → "is a"
• "stands as a" → "is a"
• "marks the" → "is the"
• "acts as the" → "is the"
• "offers a" → "has a" or "is a"
• "features a" → "has a"
• "represents a" → "is a"
• "constitutes a" → "is a"
• "ventured into politics as a candidate" → "was a candidate" / "ran for office"

The fix: When you catch yourself writing "serves as," ask whether "is" works. It almost always does. Save the fancier constructions for the rare cases where the distinction matters (e.g., something literally serving a function for something else).

---

5. Drop the Negative Parallelisms

"Not only ... but also ..." and "It's not just about X, it's about Y" are AI comfort-food constructions. They create an appearance of balanced, thoughtful analysis while often saying nothing.

What it looks like:

• "not only a work of self-representation, but a visual document of..."
• "It's not just about the beat; it's part of the aggression and atmosphere."
• "not dissolution, but what Deleuze might describe as 'becoming'"

The fix: Just say the thing. If something is two things, say both without the theatrical setup: "The portrait is a visual document of her obsessions." If the contrast genuinely matters, a simple "but" or "though" does the work without the formula.

---

6. Break the Rule of Three

LLMs default to triplets — three adjectives, three noun phrases, three examples. Used occasionally, the rule of three is a fine rhetorical device. Used compulsively, it's a fingerprint.

What it looks like:

• "global SEO professionals, marketing experts, and growth hackers"
• "keynote sessions, panel discussions, and networking opportunities"
• "bold proportions, refined dynamism, and historical reverence"

The fix: Vary your list lengths. Sometimes two items is enough. Sometimes four is better. Sometimes a single well-chosen example beats a list entirely. When you do use three, make each item carry real weight — not three vaguely overlapping ways of saying the same thing.

---

7. Stop the Elegant Variation

Repetition-penalty mechanisms cause LLMs to compulsively find synonyms for words they've already used, even when repeating the word would be clearer. A person's name becomes "the protagonist," then "the key player," then "the eponymous character."

The fix: Repeat the word. Good prose uses repetition deliberately. If you mentioned "the bridge" in the last sentence, say "the bridge" again — don't switch to "the structure" or "the crossing" or "the span" just because you already used "bridge." Elegant variation is only elegant when the variation adds meaning.

---

8. Don't Manufacture Ranges

LLMs love "from X to Y" constructions, but often the two endpoints don' form a coherent scale. A real range has identifiable middle ground: "from winter to spring" works because there's a clear continuum. A false range jams unrelated concepts into the structure for rhetorical flourish.

False range: "from the singularity of the Big Bang to the grand cosmic web, from the birth and death of stars to the enigmatic dance of dark matter"

The fix: If you can't identify a meaningful midpoint between X and Y, don't use "from X to Y." Just mention the items: "stars, dark matter, the cosmic web."

---

9. Restrain the Em Dashes

Em dashes (—) are useful punctuation. LLMs overuse them, particularly in a formulaic, sales-pitch cadence that over-emphasizes clauses. Human writers use commas, parentheses, and colons for variety.

AI pattern: "The temple — a counter-symbol of unity — stands at the border, emphasizing togetherness — and transcendent faith."

The fix: Use em dashes sparingly. One per paragraph is plenty. For most parenthetical asides, commas or actual parentheses work fine. For explanations, try a colon.

---

10. Cut "It's Important to Note"

Didactic disclaimers — "it's important to note that," "it's worth mentioning," "it should be noted that," "it's crucial to understand" — are verbal throat-clearing. If something is important, stating it makes that clear. The disclaimer adds nothing except an AI tell.

The fix: Delete the preamble. Start with the important thing itself.

Bad: "It's important to note that these regulations vary by jurisdiction." Good: "These regulations vary by jurisdiction."

---

11. Don't Summarize What You Just Said

Older LLMs added "In summary" and "In conclusion" sections compulsively. Newer ones still tend to end paragraphs or sections by restating the core idea. In most contexts, this is pure padding.

The fix: End when you're done. Trust the reader to have read the preceding paragraph. If your conclusion adds a genuinely new thought or synthesis, keep it. If it just repeats the same point in different words, cut it.

---

12. Avoid Vague Attribution

LLMs attribute opinions to unnamed authorities: "scholars note," "experts agree," "it has been described as," "many have praised." This is weasel wording. Even when a real source exists, the AI often exaggerates consensus ("several publications have cited" when there's one).

The fix: Name the source or drop the attribution. "Roger Ebert called the film X" beats "critics have praised the film for X." If you don't have a specific source, present the claim as your own analysis or reframe it as observable fact.

---

13. Don't Inflate Promotional Language

LLMs struggle to maintain a neutral tone, particularly around anything culturally or commercially significant. They slip into brochure language: "nestled within the breathtaking region," "offering visitors a fascinating glimpse," "a diverse range of experiences against the backdrop of stunning natural beauty."

The fix: Describe, don't sell. State what exists, what happened, what something does. Let the reader form their own impression. If you're writing actual marketing copy, be specific about benefits rather than stacking adjectives.

Bad: "Our revolutionary platform offers a seamless, cutting-edge experience that empowers teams to unlock their full potential."

Good: "The platform syncs files across devices in under three seconds and supports offline editing."

---

14. Don't Write "Challenges and Future Prospects"

AI-generated text often ends with a section that follows the formula: "Despite its [achievements], [subject] faces challenges including [vague list]... However, with ongoing initiatives, [subject] continues to [positive vague statement]."

This is a content-free hedge that tries to appear balanced. If challenges are worth discussing, name them specifically with evidence. If not, don' add a "challenges" section just to look thorough.

---

15. Vary Your Sentence Architecture

Beyond specific patterns, AI text has a recognizable rhythm: medium-length declarative sentences, each structured similarly, rarely interrupted by fragments, questions, very short sentences, or very long ones. Human prose breathes. It varies. Some sentences are four words. Others unspool across a paragraph, piling clause upon clause, picking up momentum like a freight train before finally, at the end, making their point.

Mix it up. Start some sentences with conjunctions. Use fragments for emphasis. Occasionally let a sentence run. The variation itself signals a human hand.

---

Quick Self-Check

Before finalizing any piece of writing, scan for these clusters:

1. Significance inflation — Can you delete "pivotal," "transformative," "broader," or "legacy" without losing meaning? Do it.
2. Trailing analysis — Do sentences end with "-ing" phrases that restate the obvious? Cut them.
3. AI vocabulary clusters — Three or more words from the AI vocab list in one paragraph? Rewrite.
4. Copula avoidance — Count your uses of "is" and "are." If they're suspiciously low, you're probably over-substituting.
5. Compulsive triplets — Every list has exactly three items? Vary it.
6. Em dash density — More than one em dash per paragraph? Switch some to commas or parentheses.
7. Preamble throat-clearing — "It's worth noting," "importantly," "it should be mentioned" — delete on sight.

---

The Meta-Principle

Every pattern above is a symptom of the same underlying tendency: the model reaching for the statistically most probable way to say something, which produces text that is simultaneously generic and overblown. The antidote is always specificity and restraint. Say the concrete thing. Use the plain word. Trust the reader. Stop when you're done.