Michael Ulin retweeted
What's worked for us: the agent never grades its own homework. Separate reviewer with fresh context, and tests that exist before the diff does.
Michael Ulin retweeted
Short spec first, then give Claude the live source before it edits. Patch what's there instead of rewriting from memory. Then have Grok or GPT critique the live build and ship v2. That loop catches more than any single prompt ever will.
Michael Ulin retweeted
what's helped me most is a second model reviewing before i merge. codex catches stuff claude code misses and the other way around. made a small tool for that, it's in my pinned post
Michael Ulin retweeted
Letting Codex or Claude Code rip without tests or specs is vibe coding with extra steps, and the software factory end of that spectrum is mostly just writing the checks first.
Michael Ulin retweeted
Is anyone focused on getting high-quality software out of coding agents?
I'm surprised how much AI coding still boils down to prompting Codex or Claude Code and letting it rip.
I wrote about 5 approaches, from vibe coding to software factories. What's actually working?
scientificfounder.substack.c…
Is anyone focused on getting high-quality software out of coding agents?
I'm surprised how much AI coding still boils down to prompting Codex or Claude Code and letting it rip.
I wrote about 5 approaches, from vibe coding to software factories. What's actually working?
scientificfounder.substack.c…
Michael Ulin retweeted
The only generation that knows how computers work is also the most relaxed about AI:
Reliability improved the week we wrote rollback steps before launch. Prompt quality mattered less than knowing exactly how to unwind a bad deploy.
whimseylabs.substack.com/p/d…
Most AI pilots do not fail on model quality. They fail when nobody decides what should happen after the answer.
whimseylabs.substack.com/p/w…
AI made software cheaper. It did not make decisions faster. The teams pulling away are the ones that cut approval delay.
whimseylabs.substack.com/p/w…
This framing matters because runway math assumes stable channels. In real life, channels wobble, so distribution depth buys time.
Elger Merlo spotlighting AppPage is timely because distribution plumbing is still where indie launches leak revenue. His short-link plus device routing example pushed us to instrument source-level redirects before our next Whimsey Labs drop.
xitter.cf/elgermerlo/status/2043…
Good reminder. Execution discipline compounds quietly, then looks obvious in hindsight when teams wonder who pulled ahead.
xitter.cf/pirwot/status/20437591…
One workflow looked great in demo, then triggered two customer escalations in 48 hours. Since then, rollout plans include rollback steps before launch.
We got better AI output the week we stopped debating prompts and started assigning failure owners before end of day.
Been testing agent stacks in real daily use. I've tried OpenClaw and NanoClaw, and Nous Research's Hermes Agent is performing best for me. Setup is easy, config is clean, and the built-in memory plus skill organization makes a huge difference.
