AI Part 5: Prompt-Driven Authoring in Practice

Theory aside — what’s it actually like to use? This post is the honest answer, including the parts I find awkward.

The drafting loop

A typical session goes like this. I describe the post I want, in a few sentences: the topic, the angle, the rough shape, sometimes a couple of points I want included. The agent drafts it. I read it. I either ask for an edit (“the third section is too abstract; ground it in a concrete example”) or I say “deploy”.

The loop is faster than typing the post myself when the post is one I would have written anyway. The agent isn’t producing better prose than I would — it’s producing prose that’s close enough to what I would have written that the bottleneck stops being keyboard speed and starts being how clearly I can describe what I want. That’s a different bottleneck, and it favours people who already think in prose.

Where it works

The thing I underestimated is how much of writing a technical series is parallel work that I was doing serially. When I drafted a ten-part Fortinet study series, I wrote the introduction, then Part 1, then Part 2, in order, because that’s how a human keeps state. With the agent, I could write the outline, then ask for all ten draft openings at once, see which ones landed, and rewrite the openings I didn’t like before any of the bodies were drafted. The same total writing happened, but the iteration was on structure, not paragraphs.

It’s also good at the parts of writing I find tedious. Frontmatter. Section titles. Excerpts. The mechanical metadata that comes with every post and that I always do badly when I do it by hand. The agent does it consistently, in the format my site expects, and never forgets a tag.

Where it’s still clumsy

There are a few patterns where I just open the editor instead.

• Surgical edits to a long file. If I want to change one sentence in the middle of an existing post, the agent has to read the whole post, re-emit it with the change, and the change goes through the file write tool. It’s accurate but feels like overkill for a one-word fix. For small fixes I open the file, make the change, and tell the agent I did so.
• Rapid back-and-forth on phrasing. If I’m not sure how I want a sentence to read, I want to try four versions in five seconds. The conversational round-trip is too slow for that. Better to land roughly and edit by hand later.
• Anything visual. Layout tweaks, spacing, the look of a card. I can describe what I want, the agent can write the CSS, but I can’t see the result without rebuilding. The feedback loop is wrong-shaped for visual iteration.

What changed about how I write

The biggest shift is that I now treat outlines as the most valuable artifact in a writing session. When I worked alone, the outline was a brief I’d discard once drafting started. With an agent, the outline is the contract — every part of the series traces back to a row in a table, and when I want to reorder or rescope, I edit the table and regenerate. The actual prose is downstream.

The second shift is that I write more posts. Not because the agent makes each post easier — though it does — but because the friction of starting a post has dropped. The hardest part of writing for me has always been the first paragraph. The agent will draft a first paragraph in fifteen seconds. Often I throw it away and write my own, but the throwing-away is itself useful: it tells me what I don’t want to say, which is often clearer than what I do.

What I’d warn anyone trying this

Three things, in order of how often they bite:

1. The agent is happy to write things you wouldn’t write. It will adopt a register, a vocabulary, a level of confidence that isn’t yours, especially on topics it has lots of training data for. You have to read drafts as an editor, not as a reader. If you read them as a reader, your voice drifts.
2. The fast feedback makes you sloppy. When a draft lands in fifteen seconds, the temptation is to deploy it in fifteen seconds. Resist that. The fact that it was cheap to produce doesn’t mean it was good. Slow down to read.
3. Long-running sessions get expensive. Each turn ingests the conversation history. A series-long session ends up paying to re-read the same context every turn. Break the work into smaller sessions. Save outlines to disk so context can be reconstructed cheaply.

The next and final post in the series is about the operational gotchas — the things that broke or behaved unexpectedly, and the small list of changes I’d make if I were starting over.