Skip to main content

Writing & Content Creation

All levels
What you'll learn
  • Go from blank page to a strong draft in your voice — fast
  • Give the model the four things it actually needs: goal, audience, format, voice sample
  • Edit in single-purpose passes instead of one vague 'improve this'
  • Repurpose one piece into five, without the flattening AI usually does
  • Never publish a fabricated fact — mark and verify every claim

AI is a phenomenal writing partner if you direct it. The goal isn't to push a button and publish — it's to compress the blank-page-to-strong-draft loop, then sharpen with your own judgment. This page is the workflow, the prompts, and the traps.

The workflow that works​

Six steps. Skip any of them and quality drops sharply.

Guided walkthrough1 of 6
  1. One paragraph. 'A 600-word blog intro for first-time founders, encouraging but not saccharine, no startup jargon. Hook in the first sentence.' Ambiguity in the brief becomes generic AI voice in the draft.

Prompts that work​

The brief-and-voice starter​

First-draft brief with voice sample

Write a {FORMAT} of {LENGTH} for {AUDIENCE}.
Purpose: {ONE-SENTENCE GOAL}.
Tone: {THREE ADJECTIVES}. Avoid: {LIST}.

Here are 300 words I wrote — match this voice. Same sentence rhythm,
same vocabulary, same restraint. Do not "improve" it toward a
generic professional voice.

<voice_sample>
{PASTE A PARAGRAPH OR TWO YOU WROTE}
</voice_sample>

Mark every factual claim, statistic, quote, or attributed source
with [verify] at the end of the sentence, so I can check it before
publishing.

The single-purpose edit​

Editing pass with visible diff

Edit this text for {ONE GOAL — pick one: clarity, concision, concreteness, rhythm, cutting jargon}.
Do not change the meaning or reorder arguments.
After the edited version, list every change as a bullet: "removed X",
"replaced Y with Z", "cut sentence about W".

<text>
{PASTE}
</text>

Ask for the change list. It forces the model to make deliberate edits rather than silently rewriting to its own taste — and it lets you veto any change you disagree with.

The outline-first for long-form​

Outline before prose

I'm writing a {2000-word essay} on {topic} for {audience}.
The argument is {one-sentence thesis}.

Before drafting, give me an outline as bullets: section headings,
one line of substance per section, plus the transition between
sections. No prose yet. Aim for {N} sections.

I'll approve or edit the outline; then you draft.

The repurposing chain​

From one piece to five formats — without flattening

Here's a published piece: <source>{PASTE}</source>.

Repurpose it into:
1. A LinkedIn post (~150 words) — one specific insight, not a summary.
2. A Twitter/X thread (5–7 tweets) — start with the strongest hook, one idea per tweet.
3. A newsletter blurb (~80 words) — the "why this matters" angle for busy readers.
4. Three email subject lines that would make me open it.

Rules:
- Do NOT summarize. Each format should stand alone.
- Preserve my voice (match the source's rhythm and vocabulary).
- Different hook for each — no reused sentences.

Voice matching — the mechanics​

"Match my voice" is worthless without a sample. What actually works:

  • Paste real work, not descriptions. "Confident but casual" is fuzzy; 400 words of your writing is not.
  • Pick a sample close to the target task. Voice shifts between formats — your newsletter voice is not your LinkedIn voice.
  • Reject flattening explicitly. Add: "Do not smooth this toward a generic professional voice. Keep the fragments, the aside asides, the unusual word choices."
  • Iterate on tone knobs. After the first draft: "Half a step more direct. Less hedging. Keep the sentence rhythm."

If a model still writes in AI-voice after this, either the sample was too short, the request was too generic, or you're on a model tuned toward the mean — try Fable 5 or Sonnet 5 which handle voice-imitation better than default GPT-family styles.

The traps that make writers reject AI​

Trap 1 — Fabricated facts, plausibly delivered

The model invents statistics, quotes, source names, book titles, and even URLs. Confidently. Every one of them.Fix: Force a [verify] marker on every factual claim. Never publish an unverified stat or quote. See Hallucinations for the four levels and how to detect each.

Trap 2 — Generic AI voice

Balanced sentences. "Not only… but also…". "It's worth noting that…". "In today's fast-paced world". You know it when you see it.Fix: Voice sample + explicit rejection of smoothing. If it slips back mid-draft, paste the offending paragraph back and say "rewrite this specific paragraph in the voice sample I gave you".

Trap 3 — Over-editing into mush

Ten small "make it better" passes flatten the piece toward the mean of the model's training data. You started with a distinctive draft and ended with a competent, unmemorable one.Fix: Single-purpose passes only. When in doubt, stop editing.

Trap 4 — The em-dash / bullet tell

Excessive em-dashes, three-bullet lists after every claim, "however" between contrasting sentences. Readers now spot this instantly and discount the piece.Fix: Explicit stylistic constraints — "max two em-dashes per section", "prose only, no bullets in this piece". Or write the framing sentences yourself and only ask the model for the middle.

Trap 5 — 'Rewrite this' loops

You keep asking "make it stronger" and the model produces different mediocre drafts because you haven't told it what "stronger" means. Rounds of this waste hours.Fix: Name the specific weakness. "The third paragraph is too abstract — give me a concrete example from a real 2020s startup." "The opening is throat-clearing — start with the punchline."

Trap 6 — Publishing without human editing

The model wrote a "good enough" draft. You skimmed and published. Two weeks later you notice the awkward transition, the misused word, the paragraph that says nothing.Fix: Print or read aloud before publishing. AI cannot hear rhythm the way you can.

Model choice for writing tasks (2026)​

TaskBest defaultWhy
Long-form essay, voice-matchedClaude Sonnet 5 or Fable 5Strongest at rhythm, restraint, and voice imitation
Fast blog draft, workmanlike proseClaude Sonnet 5 or GPT-5Fast, competent, cheap
Poetry, narrative fiction, distinctive styleFable 5, then hand-editBetter at unusual voice targets
SEO/marketing copy with keyword targetsAny frontier model + explicit constraintsAll roughly equivalent; the brief matters more than the model
Editing existing proseSonnet 5 or Opus 4Best at surgical single-purpose edits
Translation / localizationSee Porting Prompts Across ModelsVoice preservation across languages is a distinct skill

Cross-provider comparison: ChatGPT for Claude Users and Gemini for Claude Users.

Ethics and disclosure​

  • You are the editor. Judgment, taste, and accountability stay with you.
  • Disclose AI assistance where it matters — journalism, academia, contracted work with disclosure clauses. Increasingly, readers assume AI involvement in everything; being explicit about what YOU brought builds trust.
  • Do not pass off fabricated quotes or statistics as researched. This has already ended careers in 2024–2025.

Quiz — check yourself​

Check yourself

0/3
  1. You want AI to write in your voice. What actually works?
  2. Your first draft came back competent but generic. You've asked 'make it better' three times and it keeps producing different mediocre drafts. What's wrong?
  3. The draft cites 'a 2023 McKinsey study showing 68% of B2B buyers…'. You've never seen the study. Ship it?

Key takeaways​

Key takeaways
  • The brief is the biggest quality lever after voice — goal, audience, format, tone, in one paragraph
  • Outline before prose for anything longer than a few paragraphs — kill weak structure cheap
  • Edit in single-purpose passes with visible change lists; never in vague 'make it better' loops
  • A 300–500 word voice sample plus explicit anti-flattening beats every 'write like me' description
  • Force [verify] tags on every factual claim; check them against primary sources before publishing
  • AI drafts, humans edit and ship — read the final draft aloud before you hit publish

Next​