#ai

I rebuilt hush, my script that swallows successful output, and every background job now reports how many tokens it saved. The numbers aren't in yet.

Jev gives an opinion and says how confident it is. What we do with that opinion is the decision, and our code makes it. So when Jev gets it wrong, the error sits with us.

Shopify's CEO calls unreviewed AI output slop grenades. Over a long migration I generated hundreds of AI pull requests and kept the slop to myself. But careful engineers can only do so much. I think stopping slop grenades takes a process that makes them impossible.

Evals, backpressure and agent guardrails are the feedback loops we already had. The mechanism didn't change, only who reads the output.

Attention budget and attention curve describe different parts of the same problem. Budget is about how much attention there is. Curve is about how it gets distributed across the context.

AI is accelerating PR volume and cracking a review bottleneck that's been straining for years. The proposed fix — let any engineer approve anything — guts the signal four-eyes is supposed to provide. There's a better answer, and it predates AI.

There's a term for what I've been focused on: AX. Agent Experience. It's been around a while, and it's where most of my time has gone recently.

Claude Code ships with a 1M token context window. We don't need it. Disable it and your context becomes something you have to manage.

There's a quiet assumption in agentic workflows that context switching stopped being a problem. It didn't. Here's how I capture mid-session discoveries without leaving the task I'm actually in.

DDD made the case for precise language in software long before AI arrived. Now everyone's rediscovering it — just without the attribution.

A principle emerging from agentic coding: verification steps that pass should produce no output; failures should dump everything. Simple idea, but it matters more than you'd expect when an LLM is the one reading the output.

Jamon Holmgren's Night Shift workflow separates human thinking from agent execution — specs by day, autonomous implementation by night. The core insight: your time is the expensive resource, tokens are cheap. Start from a draft PR, not from zero.

Tried VoiceBox, a free open-source voice cloning app that runs locally. Cloned my voice in 30 seconds from a short audio sample — surprisingly good, even with a Scottish accent.

Claude Code has official LSP plugins for code intelligence. Easy to miss if you haven't gone looking.

Claude is assuming my codebase and dependencies are always fresh. I'm experimenting with a new CLAUDE.md decision-making heuristic. I'm trying to nudge it to check before wasting a bunch of time on debugging cycles.

Daniel Griesser's custom sub-agent workflow for context management was ahead of its time — what he hand-rolled is now being shipped as first-class tooling. The core insight still stands: context is precious, and managing it intentionally is everything.

Matt Pocock argues your codebase is the biggest influence on AI output, not your prompt. His solution is deep modules from John Ousterhout's A Philosophy of Software Design. It maps perfectly to how CQRS already works — each service boundary is a deep module with commands and queries as its interface.

Matt Pocock explains why subagents are the dominant pattern for coding agents, keeping work in the first 40-50% of the context window where models perform best.