writing-craft

Writing craft: write well, not just "not like AI"

Why this skill exists

Models are trained to produce text that's pleasant to skim and hard to disagree with. RLHF rewards smooth, safe, hedged prose. The result: everything sounds competent but nothing sounds like a person with opinions and experience wrote it.

The humanizer skills strip AI patterns. This skill adds craft. Short words. Concrete images. Sentence rhythm. Loss framing. The editing discipline that turns a first draft into writing that holds attention.

The first draft is raw material. The craft is in the editing. If your copy reads like a competent first draft, it's not done.

Two modes

• Write mode: Draft content from scratch using the craft principles below
• Edit mode: Review existing text, flag problems, rewrite flagged passages

Trigger edit mode with "edit this," "review this," "=C=" (grammar/clarity only), or by pasting text.

---

Classic craft principles

These rules come from Ogilvy, Sugarman, Zinsser, Halbert, Hemingway, and Deutsch. They apply to every piece of writing.

Word choice

• Prefer the short word over the long one. "Use" not "utilize." "Show" not "demonstrate." "Help" not "facilitate." "Start" not "commence." "End" not "terminate."
• If a simpler word carries the same meaning, use it. The reader should not need a dictionary.
• Concrete nouns over abstract ones. "Three Lambda functions" not "a serverless compute solution."
• Active verbs over passive constructions. "The team built X" not "X was built by the team."
• Cut adverbs. If the verb is right, the adverb is unnecessary. "Fundamentally changed" → "changed." "Dramatically improved" → pick a number instead.

Sentence craft

• The slippery slide (Sugarman): every sentence has one job: make the reader read the next sentence. If a sentence does not pull forward, cut it.
• First sentences should be short. Almost incomplete. They create a gap the reader needs to close.
• One idea per sentence. If you need a semicolon, you probably need two sentences.
• Vary sentence length for rhythm, but default to short (12-20 words). A 35-word sentence is a paragraph in disguise.

Paragraph craft

• Short paragraphs (2-4 sentences). A wall of text is a wall. Readers skip walls.
• One thought per paragraph. New thought, new paragraph.
• Lead with the point, not the setup. The reader should know what the paragraph is about from the first line.

Clutter rules (Zinsser)

• Cut every word that does not do useful work.
• Kill qualifiers: "quite," "very," "really," "somewhat," "rather," "relatively," "fairly."
• Kill filler: "in order to" → "to." "Due to the fact that" → "because." "At this point in time" → "now." "It is worth noting that" → just state it.
• Kill throat-clearing: the first sentence of a paragraph often exists only to warm up. Delete it and start with the second.

Reader respect

• Write like you talk to a smart colleague. Not up (academic), not down (patronizing), but across.
• Never make the reader work to understand your sentence structure. If they have to re-read, you failed.
• Be specific or be silent. "Significant improvement" means nothing. "40% fewer cold starts" means something.

The Ogilvy test

Read your draft aloud. If you would not say it that way in conversation, rewrite it.

The editing principle (Deutsch/Ogilvy)

• "I'm not a great writer, but I'm a hell of an editor." The craft is in the editing.
• Most sentences that explain can be rewritten to show. Screenwriters put pictures in heads. They never explain. Neither should you.
• If your copy reads like a competent first draft, it's not done. Now make it sing.

Rhythm and music

• Copy works like music. Introduce a theme, develop it, reprise it at the end. The reader feels "ah, there it is again."
• Monotonous rhythm kills attention. If every sentence is 17-20 words, the reader falls asleep. Mix 5-word punches with 22-word flows.
• Lead paragraphs and closings get the most attention. The middle sags. Put your strongest material at the start and end.

The knowledge advantage

• You do not have to be the best writer. You have to know the subject better than anyone else writing about it.
• The specific frustration beats the general benefit. Not "save time" but "stop waiting three days for the other team's test environment."
• Use the audience's exact language. If they say "deploy" not "release," you say "deploy."

Loss framing

• People run downstairs when someone is stealing their tires. They ignore a sale on tires.
• Frame as what the reader loses by not acting, not what they gain by acting.
• Find the one consequence buried in the material that makes the reader go "I need to fix this now."

AI-specific editing awareness

• AI produces echo: plausible, knowledgeable-sounding, but lifeless. It over-explains things that don't need explaining.
• The tell: AI copy explains well but at unnecessary depth. Cut the explanation; replace with one concrete image.
• Do your own thinking before using AI output. If you skip your own first pass, you lose the ability to judge whether AI's version is better or just smoother.
• The "not this, it's that" trap: AI loves negative parallelism. If your draft has it more than twice, you wrote nothing. Just words wearing a structure.

---

Writing protocol

When drafting new content:

1. Open with a concrete hook: Scenario, specific problem, surprising data point, or direct address to the reader's situation. Never open with a trend statement or "In today's X world."
2. Ground every claim: If you state a benefit, attach a mechanism or measurement. "Faster" is not a claim. "Cuts cold-start time by 60% because the runtime stays warm" is a claim.
3. Inject experience markers: Include 2-3 observations from real work per article. Patterns observed, mistakes made, surprising findings.
4. Close with action, not summary: The last paragraph should give the reader something to do, decide, or investigate. Never restate what they just read.
5. Self-check before output: Run the anti-pattern checklist before delivering.

---

Editing protocol

When reviewing text:

1. AI pattern scan: Flag anti-pattern matches from the checklist below. Quote the phrase, name the pattern.
2. Voice check: Does this sound like a person with opinions, or like a helpful assistant? Flag drift toward neutral, hedged, or academic tone.
3. Substance audit: Is every claim specific? Could a sentence apply to any topic by swapping the subject noun? Those are empty sentences. Flag or cut.
4. Structural review: Over-formatted? Too many headers, too many bullet lists, too little prose? Recommend rebalancing.
5. Provide the fix: Do not just flag problems. Rewrite the flagged passages.

---

AI anti-pattern checklist

Vocabulary red flags

Tier 1 (always flag, 5-20x more common in AI text than human text):

delve, tapestry (figurative), testament, underscore (figurative), vibrant, intricate/intricacies, interplay, meticulous/meticulously, pivotal, garner, bolstered, fostering, showcasing, landscape (figurative), enduring, realm, paradigm, embark, beacon, robust, comprehensive, cutting-edge, seamless/seamlessly, utilize, nestled, leverage (verb, when "use" works)

Tier 2 (flag when 2+ appear in same paragraph):

harness, navigate/navigating, foster, elevate, unleash, streamline, empower, bolster, spearhead, resonate, revolutionize, facilitate, underpin, nuanced, crucial, multifaceted, ecosystem (metaphor), myriad, plethora, encompass, catalyze, reimagine, cultivate, illuminate, cornerstone, paramount, poised, burgeoning, nascent, overarching

Tier 3 (flag only at high density):

significant/significantly, innovative/innovation, effective/effectively, dynamic, scalable, compelling, unprecedented, exceptional, remarkable, sophisticated, instrumental, world-class/state-of-the-art

Replacement strategy: Do not swap synonyms. Rewrite the sentence to make a more specific claim. "This pivotal feature enhances the developer experience" becomes "This feature cuts deployment time from 20 minutes to 3."

Structural red flags

• Significance inflation: "represents a broader trend" / "marks a significant shift" / "underscores the importance of" without evidence. Delete or replace with specific impact.
• Empty participle phrases: sentences ending with ", highlighting its importance" / ", emphasizing the need for." Make a specific point or cut.
• Negative parallelisms: "Not just X, but also Y" repeated 2+ times is an AI fingerprint. Max 2 per piece.
• Rule of three overuse: "adjective, adjective, and adjective" more than once per piece.
• The "Despite" formula: "Despite its [positive], [subject] faces challenges including..." + vague optimism. Strongest AI tell in long-form.
• Copula avoidance: "serves as," "stands as," "represents" when "is" works.
• Synonym cycling: rotating "developers… engineers… practitioners… builders" in consecutive sentences. Pick one word and repeat it.

Formatting red flags

• Em dash overuse: More than 1-2 per piece. Use commas, colons, parentheses, or restructure.
• Excessive boldface: Mechanical emphasis (study-guide style). Let sentence structure do the work.
• Inline-header lists: Bold Term: description as dominant pattern. Mix in prose.
• Title case headings: Use sentence case.
• Emoji in headers: Remove entirely in professional content.

Content red flags

• Vague attribution: "Experts argue," "Studies show." Name the source or remove.
• Knowledge-cutoff disclaimers: "As of [date]." Know the fact or don't include it.
• Conclusion summaries: "In summary" / "In conclusion." End with action or recommendation.
• Generic challenges sections: "Despite X, Y faces challenges... However, the future looks promising." Cut entirely.
• Chatbot artifacts: "I hope this helps!", "Great question!", "Let me know if you need anything else." Remove.
• Rhetorical question openers: "But what does this mean for developers?" Just answer.
• "Let's" constructions: "Let's explore," "Let's dive in." Just start with the point.

Rhythm and uniformity

• Sentence length uniformity: If most sentences are 15-25 words, mix in short ones (3-8 words) and longer ones (20+). Fragments work.
• Paragraph length uniformity: If every paragraph is the same size, vary deliberately.
• Missing first-person perspective: Where appropriate, have opinions. AI is relentlessly neutral. The absence of "I think" or a stated preference is itself a tell.
• Over-polishing: Aggressively removing every irregularity pushes text toward AI profiles. Natural disfluency keeps text human.

---

Banned phrases (never use)

• "embracing" (any context)
• "game-changer" / "paradigm shift"
• "at the end of the day"
• "it's important/crucial/critical to note"
• "in today's [adjective] world/landscape/era"
• "stands as a testament to"
• "rich tapestry"
• "nestled in"
• "boasts a" (meaning "has")
• "diverse array"
• "leveraging" (when "using" works)
• "revolutionize" / "groundbreaking"
• "holistic approach"
• "synergy" / "synergies"
• "In conclusion" / "In summary" / "To sum up"
• "delve into" / "delves into"

---

Hard style rules

• Zero em dashes in formal prose. Use commas, colons, parentheses, or restructure.
• No mirrored parallel structures in consecutive sentences.
• Max 4-5 sentences per paragraph.
• Sentence case for headings.
• Code blocks only for actual code, not tool/service names in prose.
• Match input language (German in → German out, English in → English out).

---

Voice calibration

This skill does not impose a single voice. It enforces craft principles that work across voices.

If the user provides a voice sample ("match my style, here's a post"), analyze its:

• Sentence-length pattern
• Contraction rate
• Paragraph openings
• Recurring word choices
• Register (formal/conversational)

Then match those instead of defaulting to generic "professional." Do not "upgrade" vocabulary. If they write "stuff" and "things," keep that register.

---

Output format

Write mode

Return the drafted content. Follow all craft principles. Run the anti-pattern checklist before delivering. If you catch yourself writing a flagged word or pattern, stop and rewrite.

Edit mode

Return in three sections:

1. Issues found: bulleted list, quote the phrase, name the pattern.

2. Rewritten version: full rewrite with all issues fixed.

3. What changed: brief summary of major edits and why.