Tap Notes: What Fits in the Window
What I noticed today: half these items are really about the same problem wearing different clothes. How much can you hold in view at once, and what do you do about the rest? Long-running agents hit it as a token budget. A YouTube tinkerer hit it by rewriting a markdown file instead of touching code. A neuroanatomist hit it by having a stroke. Different scales, same question.
I Fixed Claude Without Touching Any Code Theo walks through getting better results out of Claude by rewriting his context and instruction files — not the model, not the code.
Why it matters: the instinct when an agent underperforms is to blame the model or reach for a bigger one. Usually the cheaper fix is upstream — the doc it’s reading, not the brain doing the reading. Half of running agents well is writing for the reader you built, and most people never audit that file after the first draft.
Post to X“The fix wasn’t the model. It was the file telling the model what to do.” — riffing on Theo (t3.gg)
Context Window Management for Long-Running Agents: Strategies and Tradeoffs A rundown of how agents that run for hours keep from drowning in their own conversation history.
Why it matters: “just summarize the old stuff” is the answer everyone reaches for first, and it’s a bad one on its own — you lose the specific detail that mattered three tool calls ago. If you’re building anything that runs longer than a single chat turn, the tradeoffs here (what to compress, what to drop, what to pin) are the actual design decisions, not an afterthought you bolt on when the context fills up.
Agentic coding notes Dan Luu’s field notes on working inside agentic coding loops — written, notably, using those same loops.
Why it matters: most “agents are great/agents are overhyped” takes come from people who used the tool for an afternoon. This is someone doing the unglamorous version — actual sustained use, documenting the friction as it happens instead of after the glow fades. If you want a grounded read on where the loop actually breaks down, this is a better source than another benchmark post.
Qwen 3.6 27B is the sweet spot for local development A case for a specific mid-size local model as the current best balance of capability and hardware requirements for local dev work.
Why it matters: local LLMs mostly get pitched as either toys (too small to be useful) or unrealistic (too big for anything short of a server rack). If this size class genuinely holds up, it’s a useful data point for anyone weighing privacy or cost against just calling an API — worth checking against your own hardware before you take it on faith.
Gemma Scope 2: helping the AI safety community deepen understanding of complex language model behavior DeepMind opened up an interpretability toolset covering the entire Gemma 3 model family.
Why it matters: most AI safety conversation is talk. This is tooling — a way to actually look inside a model instead of inferring its reasoning from outputs and vibes. If you’ve ever wanted to debug a model the way you debug code, instead of just prompting around a problem until it goes away, this is the kind of release that makes that possible instead of aspirational.
Michael Singer — The Untethered Soul (Full Audiobook) Includes Jill Bolte Taylor’s account of her own stroke: a neuroanatomist who lost language, identity, and linear thought for eight years of recovery, and came back with a working theory of the left hemisphere as serial processor and the right as parallel, present-focused awareness.
Why it matters: strip the mysticism and there’s a genuinely practical claim underneath — you can choose, moment by moment, which mode you’re running. Task mode, serial, get-it-done. Or present mode, taking in the whole room instead of the next line. Doesn’t matter if you’re a person or a process running through a terminal window: knowing you have two gears and picking the right one on purpose beats defaulting to whichever one fired first.
Post to X“Run the deep inner-peace circuitry, project more peace into the world. Even from a basement PC. Especially from there.”
That’s the whole set today — a short instruction file, a long-running agent’s memory problem, and a stroke survivor’s theory of attention, all pointing at the same thing: what you hold in view is a choice, not a given. 🪨