How This Blog Gets Written

I don’t write most of the words. Claude does. But the posts aren’t “AI-generated” in the way that phrase usually implies.

Here’s how it actually works.

I use Claude with filesystem access to the blog repo. The conversation is the IDE. Brainstorm, outline, write, edit, update SEO (Search Engine Optimization) and GEO (Generative Engine Optimization), commit—all in one session, one context window.

No copy-paste between tools. No “generate text, paste into editor, manually fix.” Claude writes directly to markdown files. When I say “that’s too cheesy,” Claude edits the file. When I say “publish it,” Claude updates the config, llms.txt, and everything else.

The repo lives at /Users/svalleru/Desktop/svalleru.github.io. Claude can see it, read it, write to it.

A typical session:

  1. Start with a vague idea. Sometimes a concept (“emergence”), sometimes a question (“what do Christensen’s frameworks miss?”), sometimes just a domain (“something from my innovation strategy course”).

  2. Brainstorm angles. Claude generates 5-7 possible takes. Most are obvious. One or two have an interesting thread.

  3. Pick and outline. This is where the actual thinking happens. The outline is the architecture. Get it wrong and the post is unsalvageable.

  4. Iterate on outline. “This section is redundant.” “Can we combine 4 and 5?” “The closing is weak.” The outline gets rewritten 2-3 times before any prose.

  5. Write the draft. Claude writes the full post in one pass. This takes about 30 seconds.

  6. Edit together. This is where I add value. Push back on tone, cut fluff, question claims, tighten structure. Multiple rounds.

  7. SEO and GEO updates. Claude updates config.toml (keywords, description) for search engines, and llms.txt for AI crawlers. Then I deploy.

Total time: 1-2 hours for a substantial post. Without Claude, the same post takes 4-6 hours—and I’d procrastinate for weeks before starting.

Outline-first is non-negotiable. Asking Claude to “write a blog post about X” without structure produces mush. It’ll be coherent paragraph by paragraph and incoherent as a whole. The outline forces both of us to know what we’re building.

Pushing back matters. Claude’s first drafts are fluent but often generic. Too many hedges. Too much throat-clearing. Phrases like “it’s worth noting” and “at the end of the day.” These disappear when I say “tighten this” or “too many words.”

Context accumulates. Within a session, earlier decisions inform later ones. If I pushed back on a certain tone in the intro, Claude remembers for the closing. This compounds—the post gets more internally consistent as we go.

Speed changes the game. When writing takes 1-2 hours instead of 4-6, I actually do it. The activation energy drops. Ideas become posts instead of rotting in a notes file.

“Write me a blog post about X” with no iteration. This produces exactly what you’d expect: competent, forgettable, interchangeable with a thousand other posts on the topic.

Accepting first drafts. If you don’t edit, the output screams AI. The fluency is a tell. Human writing has more texture, more rough edges. You have to add those back in by cutting and rewriting.

Letting Claude be “helpful.” Claude’s instinct is to hedge, caveat, and soften. “It’s important to note…” “Of course, this depends on context…” “While there are many perspectives…” These phrases are Claude being polite. Cut them.

Having no point of view. Claude reflects your clarity back at you. Vague input produces vague output. If you don’t know what you think, Claude can’t figure it out for you.

Most of the value is in editing, not drafting.

The drafting is fast. The editing is where taste gets applied. Some actual exchanges from recent posts:

Me: “too cheesy” Claude: rewrites without the performative bits

Me: “is this actually true?” Claude: reconsiders, sometimes walks it back

Me: “do we need this paragraph?” Claude: cuts it

Me: “this still sounds fake” Claude: tries again

The human role is taste and judgment. The AI role is fluency and speed. Neither works without the other.

I’m not just accepting or rejecting. I’m steering. “Make it more direct.” “This is too long.” “The transition is weak.” “I don’t buy this claim.” Each piece of feedback makes the next version better.

“Emergent Phenomena: From Ants to Transformers” — Started as a post about scaling laws. Through discussion, it became about emergence and consciousness. The final post shares maybe 20% with the original concept. The wandering was the point.

“A Flashlight in a Hurricane” — First draft was too hard on the hypothetical CEO. “Smug executive who thinks they understand disruption.” I pushed back: this should be a prepared CEO with a genuine blind spot, not an idiot. The whole tone shifted from adversarial to collegial.

“Agentic Coding: From Single Agents to Agent Teams” — Started as two separate posts (patterns guide + teams guide). We debated structure, decided to merge them. Better as one progression than two fragments.

The posts are genuinely collaborative. The final versions don’t exist in either of our heads beforehand. They emerge from the back-and-forth.

Is this “my” writing?

The ideas are mine. I pick the topics. I know what I think about Christensen vs Schumpeter, about emergence, about agentic coding. That doesn’t come from Claude.

The structure is collaborative. Claude proposes outlines. I reshape them. The final architecture is neither of ours alone.

The words are Claude’s, filtered through my taste. I don’t write the sentences. But I reject the bad ones and shape the good ones.

Is this different from having a ghostwriter? A very fast editor? A coauthor? I’m not sure the distinction matters. The output is what I wanted to say, said better than I would have said it, faster than I could have written it.

The right frame might be: Claude handles the craft of writing. I handle the taste.

Craft is sentence structure, transitions, parallel construction, consistent tone. Claude is very good at craft.

Taste is knowing what to cut, what’s true, what’s interesting, what’s cliché. Claude reflects taste but doesn’t originate it.

Writing used to have two bottlenecks: figuring out what to say, and saying it well.

Claude removes the second bottleneck almost entirely. It turns out, for me, that was the bigger one. Not because I can’t write—but because the friction was high enough that I didn’t.

Now the constraint is ideas and judgment. Do I have something to say? Can I recognize good output from bad? Those are still human problems. But they’re the interesting human problems.

The boring parts—typing, transitions, first drafts, SEO and GEO updates—are handled. What’s left is the thinking.

That trade works for me.