Dev Notes
Dev Notes: May 25, 2026
For a few months I was burning the first ten minutes of every morning rebuilding context. What was that branch about? Why did I leave that test failing? Which approach did I rule out, and why? My work log captured the meetings and the shipped PRs, but the session itself, what I tried and where I got stuck, was gone every time I hit /clear.
So I shipped two skills for myself. /handoff writes a structured session log at the end of a working session: branch, what shipped, dead ends, open threads. /wakeup reads the latest log when I start fresh and tells me where I left off.
Now picking up yesterday's work takes about ten seconds. The skills aren't clever. The friction is just gone. I'll start a tricky task at 4pm now because tomorrow-me will pick it up cleanly instead of re-deriving the state.
That changed how I work more than any model upgrade has in the last six months.
Here's what I keep coming back to: adoption is a memory problem before it's an intelligence problem. The model can be a year out of date and still beat its competition on stickiness if it carries context across sessions. AI coding tools all run the same race on raw capability. They're losing it on continuity.
This Week on Slightly Caffeinated
E60: Turkish Coffee Finally, Multi-Step Synthesis, and AI as Enhancement
TJ walked through the multi-pass forum analysis pipeline he shipped at Luma: thread-aware batches with structured sentiment and theme extraction in pass one, then a synthesis layer that gives admins a real view of which forums are healthy. We also got into why eval systems belong as first-class citizens for any production AI feature, the growing anti-AI sentiment outside the dev bubble, and TJ's escalating robot dog plans. Turkish coffee was finally acquired. The cup itself was a letdown.
Out now wherever you get your podcasts.
What I'm Learning
Designing a technical interview loop for a role where AI fluency is one of the signals. Quick problem: how do you evaluate AI-assisted work without disadvantaging candidates who don't already pay for Claude or Cursor?
The pattern we landed on: spin up a dedicated AI workspace per candidate for the interview window, fund it with a small fixed budget, delete it after. About $50 a candidate. Levels the field. Removes "did they bring their own tool" from the signal.
Side effect we didn't plan for: the workspace setup is a forcing function. You can't half-design the interview now. You have to know exactly what access and tools each candidate gets, because you're provisioning all of it up front.
If you're hiring for AI-augmented engineering work, worth borrowing.
Dev Tool of the Week
CLI from Steph Ango, the maker of Obsidian. Strips a webpage down to its actual content and returns clean markdown. Same engine that powers the Obsidian Web Clipper. Install with npm install -g defuddle, then defuddle parse <url> --md.
The use case I keep reaching for: pulling text out of X/Twitter, Reddit, and other scraper-hostile sites where a standard fetcher either fails or returns garbage. I have a Claude Code memory rule routing all X and Reddit extraction through Defuddle by default.
If you're already in Claude Code, Steph also publishes a skills repo at kepano/obsidian-skills with a Defuddle skill ready to drop in. Clone it into your skills directory and Claude knows when to reach for it without you wiring anything up. TJ and I covered the broader workflow on E57 if you want the longer take.
Saves tokens compared to dumping raw HTML at an LLM. Saves you the cleanup pass.
That's it for this week. Hit reply if you've got a handoff pattern that's working, or if you've tried something different for AI-assisted interview design.
-Chris