Here's an uncomfortable thing I want to get out of the way: Claude wrote the first draft of a lot of our landing pages. Not all of them. And not the final version of any of them. But the structure, the feature descriptions, the FAQ — first pass was Claude. And until I mentioned it in a couple of conversations with friends, nobody noticed.
I've been chewing on what that means. Because the easy reaction is either "AI can write now, we don't need writers" (overstated) or "using AI to write is dishonest" (overstated the other way). The truth is more interesting and more useful.
What Claude Is Good At
Given a spec.json file with a product's features, pricing, and target user, Claude produces a reasonable first draft of a landing page in about 30 seconds. The draft will include:
- A hero headline that's grammatically correct and on-topic.
- Six feature blocks that accurately describe what each feature does.
- A quickstart section that correctly reflects the tech stack.
- A pricing table.
- An FAQ section with reasonable questions.
All of those are fine. None of them are bad. In the sense of "will a random visitor understand what this product does," the draft succeeds.
What Claude Is Bad At
Claude's default voice is SaaS house style. "Empower your workflow." "Unlock the full potential." "Transform your business." When I see those phrases in a draft, I delete them and rewrite. Every time. They're cheap, they signal nothing, and they make the page indistinguishable from every other SaaS page ever written.
Claude also defaults to three-adjective stacking. "Powerful, flexible, intuitive." "Modern, beautiful, fast." Those adjectives don't mean anything when they're piled on top of each other. I strip them.
And Claude, without specific guidance, produces copy that's agreeable. Everything is great. Every feature is exciting. The app does everything well. No tradeoffs, no opinions, no texture. Landing pages written that way don't sound like they were written by a person who built the thing, because they weren't. They sound like they were written by an intern who read the feature list.
The Fix Is the Voice Prompt
The difference between a Claude-draft that I have to rewrite entirely and one that I can lightly edit is the prompt. Specifically, the part of the prompt that tells Claude about the voice it should be writing in.
For Dangercorn, my voice prompt is something like:
Tim Wiley is a Pacific Northwest engineer in his 50s who has spent 30 years in IT. He writes in first-person, is dry, is a bit grumpy about enterprise SaaS, and says things like "we built this because we were tired of X." He doesn't use 'leverage' or 'unlock.' He is specific about tech (Flask, SQLite, port numbers) rather than generic ('AI-powered platform'). He's a little self-deprecating. He's willing to admit things don't work or are still being figured out.
With that context, Claude's drafts get dramatically better. Still not final-draft quality, but 70% of the way there instead of 30%.
What I Rewrite by Hand
Roughly: every hero headline. Every section headline. Every transition. Every joke. Every opinion. Every "we tried this and it didn't work" moment. Every specific technical detail that mentions a real port number or a real tool name.
What I keep: the feature descriptions (often with minor edits), the FAQ, the pricing table, the structure of the page.
In practice, a landing page ends up about 50-60% my hand and 40-50% Claude's draft, but the parts that are me are the parts that give the page its character — headlines, transitions, opinions — and the parts that are Claude are the mechanical explanatory parts where doing it by hand wouldn't have changed much.
The Meta Question
Is it dishonest to use Claude to draft copy if I don't disclose it on every page?
My answer, for now: probably not, if the final copy is mine. Every novelist uses a word processor. Every screenwriter uses Final Draft. Using Claude to produce a first draft that I heavily edit is closer to "using a word processor" than "having a ghostwriter." The final product is what I would have written, just produced faster.
But I'll be honest about where the line might be. If I took Claude's raw output and shipped it without editing, that'd be different. If I had Claude write the thesis posts (like this one about building 220 apps) and passed them off as my own thinking, that'd be different. The difference is whether the substance of the ideas came from me.
The substance of a product landing page is: here are the features, here's the pricing, here's how to install it. The substance of an opinion post is the opinion. Claude is fine for the first. I don't use it for the second, except as a second pair of eyes.
What About This Blog?
Relevant question, yes. The existing blog posts — the fleet post, the prediction markets post, the cold email retrospective — I wrote by hand. Claude didn't touch them until I pasted in a finished draft and asked for typo corrections, which is a different use case.
The blog series you're reading right now is somewhere in between. For each post, I sketched the outline and the key points, then Claude drafted the body, then I rewrote for voice and added specific details that only I would know. Claude doesn't know, for example, that the T3600's SSD has 79,000 hours on it. I do. That kind of detail has to come from the person with the lived experience.
The Claude-drafted version is never the final version. The hand-edited version is. The question is just whether the draft saves enough time to be worth the edit pass. For explanatory content, usually yes. For voice-heavy content, usually no.
The Practical Upshot
If you're building a portfolio of products like we are, Claude is part of your toolchain. Same category as your code editor or your deploy script. Pretending otherwise would be weird, and it'd also leave productivity on the table.
The skill that matters now isn't writing from scratch. It's editing. Knowing what a good landing page sounds like, being able to look at a Claude draft and instantly spot the six sentences that need to go, being willing to rewrite a paragraph by hand because the generated version is technically correct but emotionally flat.
That skill is learnable. It also doesn't get replaced by better models, because the judgment "this sentence sounds AI-generated" is itself a human skill that improves with reps. Which means the division of labor — Claude produces the scaffolding, I produce the character — looks stable to me for as long as I want to keep writing.
And the nobody-noticed part? I think that's because the hand-editing works. The pages don't read as AI copy because they aren't, anymore, by the time they're live. They just took less time to produce than they would have a few years ago.