Dev Notes
Dev Notes: June 1, 2026
I let Claude write most of a feature last week, then caught myself scrolling past the diff. Approve, move on. The code worked. I had not actually read it.
That's the part nobody warns you about. AI makes the typing fast. It also makes skimming feel safe, because the output usually runs. The cost shows up later, when something breaks in a part of the system you "wrote" but never understood.
Here's the position I keep landing on: as the model handles more of the writing, the scarce skill flips to reading. Not reading to rubber-stamp. Reading to interrogate. What did it assume? What did it skip? What surprised me here?
I've started treating the diff like a code review I owe myself. Line by line. Note what I didn't expect. The small in-flight moments, the "huh, why did it do that," are where the actual learning lives now. Skip them and you ship code you can't reason about a month later.
Claude Code builds a lever for this right into its settings. Switch the output style off Default: Explanatory walks you through why it made each implementation choice and how it fits the surrounding code, and Learning goes further, pausing to hand you small pieces to write yourself. Both trade a little speed for understanding, the exact thing you lose when you let it run unattended.
Agent capability gets all the attention. Every release is faster, cheaper, smarter. But your capability as the operator is the half of the loop you control, and it's the half that quietly degrades when you let the tool carry you.
The teams that win the next year won't be the ones with the best model. Everyone has the best model. They'll be the ones who stayed sharp enough to tell working from right.
This Week on Slightly Caffeinated
E61: Meeting Voice Review, Assembly AI, and Marketplace Plugins
TJ is building an org-wide Claude Code plugin marketplace at work: purpose-built plugins for web, mobile, and QA, each carrying its own skills and rules. That kicked off a good back-and-forth on skill context bloat, the new setting that caps how much context your skills can eat, and why TJ leans on rules over agents. Rules only load when you touch a matching file, so they cost you nothing until you need them.
Out now wherever you get your podcasts.
What I'm Learning
I built a skill that reviews how I talk in meetings. I saw a video about using a Zoom integration for speaking analytics, pointed Claude at the video transcript to pull the ideas, then had it build the same thing against the meeting transcripts already in my notes.
Give it a time window and it finds my lines across every transcript, then reports my filler words, hedges, and longest monologues in an HTML dashboard with trends. The results were humbling. "Just," "kind of," "I think," "I feel like," on repeat.
What it can't do is tone, pacing, or pauses. That needs the audio, and I don't save it. Transcript-only gets you the verbal tics, not the delivery.
The takeaway: you probably already have the raw material for a feedback loop you've never run. Meeting transcripts, PR comments, old standup notes. Point an AI at the data you already generate and ask what pattern it sees.
Dev Tool of the Week
A speech-to-text API that actually handles speaker diarization well.
I swapped it into my podcast transcription workflow to replace OpenAI's Whisper, which never tagged who was speaking. AssemblyAI does, and it infers names from context. When one host says "TJ, what do you think," it tags that speaker as TJ for the rest of the transcript. No manual cleanup.
The free tier starts with $50 of credit. At a couple of podcasts a week, that lasts a long time.
If you're transcribing anything with more than one voice, the diarization alone is worth the swap.
That's it for this week. Hit reply and tell me what your AI diff-review habit looks like, or if you've found a transcription tool you actually trust.
-Chris