Behind the posts

I write my posts with Claude: I dictate, Claude drafts, and I review until it sounds like me. After each post, Claude logs every correction: what went wrong, and who caught it. This page collects those numbers, post by post.

241 corrections so far, 208 of them caught by me.

What the numbers show so far

Claude catches its own mistakes when it checks something: a name, a screenshot, the page as built. Most of the 33 it caught were misheard words from my dictation, and things that only looked wrong on the page.

The rest I catch myself: sentences that say less than I mean, words I would never use, and details that sound plausible but are not true.

Agentic shopping: what I learned before buying anything

18 October 2026

93 corrections, 85 of them caught by me and 8 by Claude.

  • Clarity36
  • Not sounding like me17
  • Wrong, but plausible17
  • Not landing with my readers9
  • Diagrams11
  • Not checked on the page1
  • Dictation1
  • Process1
  • Clarity. 36 of the 93. Paragraphs that went too fast, a sentence that only made sense two sections later, and a heading like "What I thought would stop it" that said nothing.
  • Not sounding like me. 17. "Nappies", a version of me that jumped in and only then found the risks, and a payment card for our agents where the excitement took over.
  • Wrong, but plausible. 17. A ChatGPT that could browse the web from day one, notifications that would warn me about a changed address (they only cover orders), and an agent that supposedly made no decisions on this website, when it had put changes live without asking.
  • Not landing with my readers. 9. Sending agents to Albert Heijn, two days after its cybersecurity incident, "thinking like a hacker" and a "kill switch", and links to Claude's own product pages.
  • Diagrams. 11. A diaper price that looked wrong because the brand was missing, a sketch of Google around 2000 with shops that did not exist yet, arrows that went one way where a chat goes both ways, and arrows running through labels. Claude caught five of them.
  • The rest. A code example that scrolled sideways on a phone, one misheard word ("Macro" for Makro), and a diagram described in text instead of drawn.

Claude caught its own mistakes when it checked something: a screenshot, its own documentation. I caught the rest, and seven of those were my own dictated words, which only looked wrong once they were on the page.

The contextual layer, parts 1 and 2

2 October 2026

103 corrections, 90 of them caught by me and 13 by Claude.

  • Clarity48
  • Not sounding like me20
  • Wrong, but plausible13
  • Not landing with my readers9
  • Diagrams4
  • Not checked on the page3
  • Dictation6
  • Clarity. 48 of the 103. Sentences that said less than I meant, like an agent that "follows" the paths in the graph where the layer makes it follow them, and edits bolted on instead of worked in.
  • Not sounding like me. 20. Claims stronger than I would make, like "the four steps to build one", and words I do not use, like "checkable" or "a lovely cooking shop".
  • Wrong, but plausible. 13. A model that retrained every night, a team on holiday in the busiest week of the year, and predictions rerun for everyone when the fix was for consumers only.
  • Not landing with my readers. 9. A first example about groceries that any model already understands, a title that said nothing about the subject, and "data warehouse" where data scientists use a data platform.
  • Dictation. 6 words Wispr Flow misheard, like "Gardner" for Gartner. Claude caught every one of them.
  • Diagrams. 4. Arrows that read backwards, like "team is maintained by dataset". Claude caught one of them from a screenshot.
  • Not checked on the page. 3. A book link that broke on the live page, and photo crops that cut off a shop sign and two passers-by at the head. Claude found the broken link itself.

Claude caught its own mistakes when it checked something: a name, a screenshot, the page as built. The errors that mattered most read well and were wrong, and it took my own experience with these systems to see it.

Why I stopped writing, and why I started again.

24 September 2026

45 corrections, 33 of them caught by me and 12 by Claude.

  • Clarity9
  • Not sounding like me19
  • Wrong, but plausible4
  • Not landing with my readers3
  • Not checked on the page2
  • Dictation8
  • Not sounding like me. 19 of the 45. Drama I would never use ("it knocked the wind out of me"), filler like "which is great", and an ending that made me sound as if I had given up.
  • Flow. 9. A paragraph pointing back at nothing, and the same transition used twice.
  • Dictation. 8 names and words Wispr Flow misheard, like "clampdown" for Calmcode. Claude caught every one of them.
  • Invented or unchecked. 4, like my kids playing hide and seek for "an hour". They read fine, which is exactly the problem.
  • The rest. Jargon and links that did not suit my readers, and two styling mistakes that only a screenshot showed.

Claude caught its own mistakes when it checked something: a name, a screenshot, a count. The ones about my voice and my life, only I could catch. So read everything before you publish, especially the parts that sound plausible.

The log is kept by a Claude Code skill. If you write with AI as well, you are welcome to read and reuse the meta-learnings skill.