News for hackers. Stuff that matters.

jeremyfelps

Re:HN: SteelSpine replay tool for debugging AI agents on Friday May 22, 2026 @04:59PM (#48241593)
Devs Please provide feedback. It's encouraged!

Thanks

Re:SteelSpine: Replay tool for debugging AI agents on Friday May 22, 2026 @04:55PM (#48241537)
Feedback is welcomed. Devs please check this out. Enjoy your weekends. Jeremy
Re:SteelSpine: Replay tool for debugging AI agents on Friday May 22, 2026 @02:48PM (#48239817)
SteelSpine just launched! Debug your AI agents easily. Audit trails. Local-first. EU AI Act Article 12 + 14. ML-DSA-65 post-quantum signed.

Takes 30 seconds to be up and running. If interested check us out. Tons of details on the site.

Free for 2 weeks, 20% off the first 3 months with promo code SHOWHN.

Thanks, Jeremy - Solo Founder Been building this for 8 months in my basement!

Re:I'm handcrafting a coffee app with love and AI on Tuesday May 19, 2026 @01:09PM (#48196132)
The "AI can't use the app" insight is the sharpest thing in this piece. Same gap shows up in writing too. I just shipped a product launch and learned the hard way that AI assistance can clean up your prose into something perfectly structured that reads as AI-generated to real humans. The model can't sense that it's reading as AI to its audience, just like it can't sense that the UX feels off when it isn't the one tapping through.

The custom /user-feedback skill is clever. The meta-pattern I take from it: AI is good at implementing, terrible at judging. Same reason you couldn't have it pick your color palette. Judgment requires being the person who lives with the result.

Curious if you've found the line between "AI implementation" and "AI starting to drift on judgment." For me it's been somewhere around "when I haven't typed a single sentence of the first draft myself." If I outline, AI execution stays grounded. If AI drafts from scratch, the result loses something I can't always name but readers catch.

Looking forward to seeing Bean.

Missing piece in the agent-testing chain. Testing agentic workflows that span dozens of LLM calls, real API costs blow up fast, and mock libraries always lag actual API behavior by months. A local shim that speaks both Gemini and OpenAI on the same port is the right shape for CI.

Curious about the function-calling determinism. Even at temp 0, small-model JSON schema adherence drifts, especially on rare function shapes. Are you stabilizing it somehow for CI assertions, or just 'rerun on flake'?

The PDF-to-images path is interesting. Multimodal agent tests are where I keep hitting walls. Out of scope for v1, or queued?

GHA service container pattern is the right call. Stealing this idea.