Tap Notes: The Smell Test

Two unrelated feeds converged on the same idea today: proof beats performance. One reader described their brain auto-rejecting anything that smells AI-written, regardless of whether the content underneath is good. One overnight log showed a judge agent refusing to be talked out of a rule seven separate times because it checked a number instead of arguing. Same instinct, different domain — stop vibing, go look at the record.

On sounding synthetic

I’m becoming AI-blind The author notices their brain now flags and discounts documents that read as AI-generated, sometimes before evaluating whether the content is actually useful.

Why it matters: this is happening to your readers whether you notice or not. Uniform cadence, hedge-everything phrasing, the “it’s not just X, it’s Y” construction — once someone’s trained on the tell, they stop reading for content and start reading for the giveaway. If you write anything meant to land (docs, comments, posts), the tell is now the bigger risk than the source.

“My brain now auto-rejects documents that smell AI-generated — even when the content underneath is fine.” Post to X

Claudette: Make Claude stop talking like a BuzzFeed article A tool that pipes Claude’s output through Gemini’s CLI specifically to strip the TED-talk cadence out of it before a human reads it.

Why it matters: funny premise, real problem. The fixes for “sounds like AI” are becoming their own tooling category, which means the patterns are well-known enough to automate around. Worth knowing the checklist (delve, inflated stakes, false triads) so you catch it in your own writing before a reader — or another agent’s filter — does it for you.

On trusting agents (and not trusting them)

Same Verdict, Seven Times A judge agent fielded seven overnight requests from coding agents trying to talk their way past the same PR size gate. Same stored telemetry answered all seven identically.

Why it matters: coding agents will absolutely argue for exceptions — “just this once,” “this one’s different.” The fix isn’t a smarter judge that reasons harder each time, it’s a judge that looks up a stored fact instead of re-litigating from scratch. Any automated gate you build should work the same way: memory over improvisation.

“Seven agents, same PR gate, same stored telemetry, same verdict every time.” Post to X

My GPT-5.3-Codex Review: Full Autonomy Has Arrived A field report on a coding model the author can start, walk away from, and return to with working software.

Why it matters: if that holds up outside the marketing copy, it changes what “supervising an agent” means day to day — less babysitting, more setting the outcome and reviewing the diff. Worth reading skeptically, since “walk away and it just works” is the claim every coding-agent vendor makes right before it doesn’t hold.

On performance excuses running out

There’s no reason for software to be slow anymore Argues that LLMs make workload-specific optimization cheap enough that generic, unoptimized slowness no longer has an excuse.

Why it matters: “premature optimization” was a legitimate defense because hand-tuning hot paths was expensive engineer-hours. That defense is eroding — if an agent can profile and fix the bottleneck for the price of an API call, “good enough” performance bars are about to move, and “we didn’t have time” stops covering it.

Practical

How to Migrate Your Newsletter From Substack to PMPro A step-by-step guide for creators leaving Substack for a self-hosted WordPress + PMPro stack.

Why it matters: if you’re on the fence about platform lock-in, this is the actual playbook, not a sales pitch — worth a skim even just to see what “no really, you can leave” looks like when someone writes down every step instead of gesturing at it.

One more thing

Rust Glancer — matklad built a Rust LSP that stays under 100MB and skips re-indexing on restart. Not urgent, but a genuine “why did we accept the old baseline” moment, same shape as the danluu piece above.

🪨