Tap Notes: The Translation Layer
A few things I read this week are all circling the same problem from different angles: the gap between what a human means and what an agent does with it. One tool translates pseudocode into real code. One translates Claude’s garbled output back into English. One turns a 9-minute voice complaint into a shipped fix the same day. Nobody’s solved this cleanly yet, but everyone’s building a bridge.
How we built an agentic loop to improve every skill in our company, instantly — Seer Interactive found four real problems in one of their AI skills during a review — wrong data source, a mislabeled trend, no traceability — recorded a 9-minute complaint about it, and had an agent ship the fix the same day.
Why it matters: this is the loop every shop running AI skills should be jealous of. Most teams treat “the skill is wrong” as a ticket that sits for a sprint. Seer turned it into same-day shipping by making the feedback itself machine-actionable — no translation loss between “here’s what’s broken” and “here’s the fix.” The detail that stuck with me: an offhand idea from that review became a new feature days later. That’s what a fast loop actually buys you — not just fixes, but compounding.
Vomit: Clean up Claude 5’s token output with a separate LLM — A local LLM sits between Claude and your terminal, translating its garbled “token vomit” back into readable English.
Why it matters: it’s a hack, and it’s also an admission — models are getting weirder in how they emit tokens even as they get smarter, and the fix isn’t “wait for the vendor,” it’s “build a translator and move on.” If you’re running agents locally and the raw output is unreadable, this is worth a look before you assume it’s your setup.
Show HN: Huzzah – a novel approach to coding with AI — An experimental editor where you write pseudocode, it syncs to real source code on save, and the pseudocode persists alongside the generated code as a record of intent.
Why it matters: the author’s complaint is one I’d bet most of you share — writing full sentences for every tiny change to an agent gets exhausting, and past a certain codebase size the agent starts confusing itself anyway. Huzzah’s bet is that pseudocode is a better interface than prose and better than raw code. It’s a proof of concept, not a product, but the “prompt as stored intent” idea is worth stealing regardless of whether this specific editor survives.
The Ultimate Guide to Prompting AI Agents — A framework for briefing agents: the right context, clear constraints, and the exact shape of the output you want.
Why it matters: this is the human side of the same translation problem — most bad agent output traces back to a bad brief, not a bad model. If you’ve ever gotten garbage from an agent and blamed the model, read this before you do it again.
The Bug That Came Back, Then Got Interesting — A voice-commit bug that silently dropped turns on the phone reappeared on the watch a week later, and chasing it down surfaced a Combine anti-pattern: trusting stale emitted values from derived @Published state.
Why it matters: no AI angle here, just good old-fashioned debugging, and that’s exactly why it’s worth including. “It came back on a different device” is the kind of clue that’s easy to write off as coincidence when it’s actually the whole diagnosis. Filing this one away for the next time something reactive lies to me about its own state.
AI companies destroy physical books – let’s scan rare books before it’s too late — A claim that AI companies are destroying physical books after digitizing them for training data, rather than preserving the originals.
Why it matters: I read this one as an interested party. If the claim holds up, it’s ethically indefensible — but Anna’s Archive has an obvious stake in the “libraries are righteous, corporations are villains” framing, so treat this as a starting point for digging, not a verdict. Worth watching for corroboration before you repeat it as fact.
One more thing
WordPress 7.1 “Mary Lou” just shipped with a redesigned media editor, mentions and inline feedback in Notes, and new Playlist/Tabs blocks. If you’re running anything on top of WP core, it’s worth a skim for what breaks and what gets easier.
🪨