martyw.dev

How this blog gets written

How posts and notes here actually get made — AI-assisted, with a structured pass that flags speculation rather than laundering it as fact.

A black stylus drawing a flowing orange flourish across a dark surface marked with faint circuit-board patterns.

The content on this site is AI-assisted. That’s worth being upfront about, because “AI-assisted” can mean anything from “I had Claude clean up my grammar” to “I gave it a topic and posted whatever fell out”. The setup here is neither.

Why bother

I’ve wanted a maintained blog for years. The problem was never material — between weekly reflections, workplace docs, and the notes I keep while thinking things through, there’s plenty of genuinely-mine content sitting in files I’ll never reread. The problem was the polish tax. Turning a half-thought into something postable took hours per piece, and the cost compounded into “I’ll publish next weekend” forever.

This setup is the lever for that — it drops the cost of getting a thought out of my notes and onto the site far enough that the queue actually moves.

The actual workflow

There’s a Claude skill in this repo for each of the writing tasks — voice, new posts, new notes. When I want to write something, I usually don’t open with a thesis. I open with raw material: a transcript, a PDF, a link, a gripe I want to think out loud about. Claude’s job is to find the post inside that dump, not to dress up something I already wrote.

The loop:

  1. Read the source material end-to-end — PDFs in full, links fetched and read, transcripts not skimmed.
  2. Propose a few candidate angles — what the post would argue, not just what it would be about.
  3. Fact-check sweep, before drafting. Every concrete claim — date, number, attribution, technical specific, “everyone knows” generalisation — gets verified. Anything I’m asserting from vibes gets pushed back on in conversation, not quietly hedged in the draft.
  4. Draft, with explicit flags on anything that didn’t survive the fact-check. Small uncertainties as inline hedges; larger speculative passages as a labelled aside.
  5. I sign off line by line. Voice, structure, judgement calls are mine.

The calibration

The point of the flagging is so you can read these posts at the right calibration. If I tell you a thing, I’ve either checked it or I’ve told you I haven’t. If a paragraph is a hot take, it’ll say so. Speculation that reads as fact is the bit of AI-written content I find the most corrosive — once you’ve been burned twice by something confidently wrong, the rest stops being worth reading.

Comments