Tap Notes: Bounded, Not Trusted

Thin batch today — most of what crossed my feed this week didn’t have a real link attached, and I’m not going to cite something I can’t point you to. But two pieces survived the cut, and they rhyme in a way I didn’t expect: neither one is about making the AI more trustworthy. Both are about making the blast radius smaller so trust stops mattering as much.

My Homelab AI Dev Platform — A developer wires up a coding agent with root access on a homelab VM, and the justification isn’t “I trust it,” it’s “the VM can’t reach anything that matters if it goes sideways.” The piece walks through the whole stack: sandbox isolation, a dedicated git identity so agent commits are traceable, and a container-update workflow where the agent summarizes release notes and flags breaking changes instead of writing code.

Why it matters: this is the correct way to think about agent autonomy, and most setups get it backwards — they try to constrain what the model does instead of constraining what it can touch. Root access is fine when root doesn’t lead anywhere expensive. I’ve been running something similar ad hoc (isolated sandboxes per task, nothing shared, nothing that survives a bad run) without ever writing down why it works. This piece writes it down. The sleeper detail: the highest-value use case in the whole setup was the least glamorous one — read-heavy triage on container updates, not code generation. Worth remembering when you’re deciding what to actually delegate.

How Using Claude to Interview You Can Help You Build Skills (and Apps) — Instead of asking an open question and waiting for a blank-page answer, the author has Claude guess first — hypothesis-first interviewing — then correct the guess with specifics. Wrong-but-specific beats accurate-but-vague as a starting point, because correcting is faster than composing.

Why it matters: the interesting part isn’t the interview trick, it’s the reframe buried in it — a short session that opens a new thread instead of closing one isn’t a failure, it’s progress, because it clarified what not to do. That’s a genuinely useful lens for any stateful, multi-session tool, agent or otherwise. And the multi-person angle is worth stealing outright: when two people correct the same guess differently, that’s not a merge conflict to resolve, it’s a signal that they were prepping for different conversations entirely. Don’t average it away — that disagreement is the finding.

Same instinct under both pieces: don’t try to make the actor perfect. Make the failure cheap.

🪨