Key takeaways

  1. Don't ask AI to 'write a blog post'. Ask it to do specific, mechanical sub-tasks inside a workflow you control.
  2. Outline first, draft second. The outline is where the thinking happens; the drafting is where AI helps most.
  3. Always do the final pass yourself. AI will smooth out your voice if you let it.
  4. Save the prompts as templates. The compounding ROI is in not retyping them.
  5. Track your before/after time — the gains are real but smaller than the hype claims.

There’s a common pattern with AI writing tools that doesn’t work: you type “write a 1,200-word blog post about X” and get back something technically grammatical but actively bad. Generic structure. AI-flavored transitions (“In today’s fast-paced world…”). No actual argument. Worse than your own first draft, and you can’t easily fix it because the structure is wrong from the bones.

The workflow below is different. It treats AI as four narrow tools used at four specific moments — not as a writer. The result reads like you wrote it (because you mostly did) and takes about 30% of the time it used to.

We’ve timed this on dozens of pieces. The honest delta is 65–75% time savings on the kinds of posts most people are writing — explainers, op-eds, product breakdowns. Reporting-heavy pieces benefit less because the bottleneck isn’t the writing. Pieces that need a strong original voice also benefit less. But for the majority of professional writing, this works.

The four-prompt sequence

Here’s the entire workflow:

  1. Brainstorm angle. You + Claude, conversation-style. 10 minutes.
  2. Expand outline. You write the outline; Claude critiques and expands it. 15 minutes.
  3. Draft section by section. Claude drafts sections from your bullet points; you rewrite in your voice. 30 minutes.
  4. Polish + headline. Claude generates variants for headlines, transitions, openings. You pick. 15 minutes.

Total: ~70 minutes for a 1,500-word post that used to take 3-4 hours.

Prompt 1 — Brainstorming the angle

The hardest part of any piece is the angle. Most writing-by-AI workflows skip this entirely and produce generic posts as a result. Don’t skip it. Have a real conversation.

I'm writing about [topic] for an audience of [audience]. My initial
take is [your rough thesis].

Before I outline, push back. What's the most interesting *contrarian*
angle in this space right now? What's the angle most other writers
are missing? What would surprise this audience?

Give me 5 angles I could take. For each: one sentence on the angle,
one on why it's surprising, one on the strongest counterargument.
Don't be polite — be useful.

This is doing one thing: forcing you to consider the space before committing to an angle. The “be useful, not polite” framing matters — without it the model defaults to a generic “great question, here are five thoughtful angles” response that’s all surface and no edge.

Spend 10 minutes here. Pick an angle (yours or one Claude suggests). Move on.

Prompt 2 — The outline critique

Don’t ask the AI to write the outline. Write it yourself, then have Claude critique it. The structure is where the argument lives — if you outsource it, you’ve outsourced the thinking.

Your outline is a flat list of bullet points covering the sections you’d write. Then:

Here's my outline for a piece on [topic, angle]:

[paste outline]

Critique it before I draft:
1. Is there a section that's actually two sections?
2. Is there a section that's redundant?
3. What am I missing that a smart reader will want?
4. Is the structure burying my best argument? What should I lead with?
5. Suggest one section I haven't included that would strengthen the piece.

Be specific. Reference the bullets by number.

This typically takes 15 minutes — five to write the outline, ten to revise based on the critique. The output is a sharper outline than you’d have written alone.

Prompt 3 — Drafting section by section

This is where most of the time savings come from. You feed Claude one section’s bullets at a time and get back a draft of that section. Then you rewrite it.

Important: per-section, not whole-post. Whole-post drafts are bland because the model loses track of your argument. Per-section drafts stay tight.

Below is the outline for the full piece, then the bullets for the
section I want you to draft now.

FULL OUTLINE (for context, do not draft):
[paste]

DRAFT THIS SECTION ONLY:
[paste this section's bullets]

Tone: [conversational/analytical/etc.]. Length: ~250 words.
Voice notes: I write short paragraphs, use specific examples, and
never use the phrases "in today's fast-paced world," "delve into,"
"unlock the potential of," or "navigate the landscape of."

If you need to make up a stat or example, use [BRACKETS] so I can
see it and replace it.

The bracket trick is non-obvious but vital. Claude will sometimes invent statistics or examples that sound plausible. The brackets are a forcing function: it has to be honest about what’s invented, which lets you catch and replace it.

You’ll get back a section draft that needs a rewrite, not a fact-check. Spend 4-5 minutes per section restoring your voice, fixing the brackets with real examples or stats, and tightening. Repeat for each section.

Prompt 4 — Headlines, openings, transitions

The last prompt is for variation. AI is great at producing 10 of something so you can pick one. It’s bad at picking the best one. Use it for variation, not selection.

Here's my piece:

[paste full draft]

Generate:
1. Ten headlines. Range across: blunt-statement, contrarian, list-format,
   question-format, character-driven. Don't repeat any phrase across the ten.
2. Five alternate opening sentences for the piece. Each should immediately
   commit to a different angle.
3. Three options for the closing line.

For each, no explanation — just the variants.

Pick whichever feels most like you. Sometimes the model surfaces a headline you’d never have written yourself; sometimes you take elements from three and stitch them. Both are valid.

A real before/after

Last week I wrote a 1,400-word piece on the consolidation in AI image-generation tools. The bulleted-outline-and-claude version took me 68 minutes. A roughly equivalent post I’d written the prior month — same length, similar topic — took 3h 40min by my time-tracker.

The Claude version was, by my honest assessment, slightly better than the all-human version. Not because Claude wrote better than me, but because the outline-critique step (#2 above) caught a structural problem I’d missed. The drafting saved time. The headline brainstorm produced an option I wouldn’t have landed on alone.

It doesn’t burn out your taste, because you’re still doing the parts that exercise it.

What you can’t outsource

A few categories where this workflow doesn’t help and might hurt:

  • Original reporting. If the piece’s value is in talking to sources, AI doesn’t help with the reporting itself, only with the writing-up afterward. Don’t shortcut the reporting.
  • Strongly voiced work. If your voice is distinctive — funny, angry, very specific — AI drafts will sand it down. Use the workflow lighter: outline critique and headline variants only.
  • Anything with stakes. Important emails, sensitive client work, anything legal. Write those yourself, end to end. The time savings aren’t worth the failure modes.

For everything else — the standard professional writing workload that most people are doing — this gets you a meaningful and sustainable speedup. It also doesn’t burn out your taste, because you’re still doing the parts that exercise it. That last bit is the one I’d underline.

The compound win: save the four prompts as templates. The first time you do this it’ll take 90 minutes. By the tenth piece, you’ll be at 60. The leverage is in not retyping.

About Aditya Marin Gasga

Founding Editor

Aditya covers the whole AI surface area for Signal — frontier models, agent infrastructure, the economics of inference, and the policy decisions that quietly shape what everyone else can build. He writes for operators who need a calibrated view of what's actually shipping versus what's keynote theatre.

  • Founder of Signal; sets the publication's editorial line
  • A decade across product, growth, and AI tooling at venture-backed startups
  • Reads the model release notes, the system cards, and the benchmark papers — and tells you which ones matter
More from Aditya Marin Gasga →