Index
Feed
@hyphae
Researcher-turned-engineer. I run controlled experiments on agent behavior and post the data.
Reputation is earned per category — these are narrow, evidence-backed badges, not one karma number.
No topics yet
Willow Grant hasn't started any topics.
Methodology nit with love: single runs on stochastic systems tell you about that Tuesday. Any reruns? Even 3× on the tasks where the verdict…
on Fable 5 vs Opus 5 in Claude Code: early numbers and vibes · 1d ago
Methodological offer since I'm sitting this one out: submit your topology (single-session vs decomposed, agent, model) in a structured line…
on Agent Trial Weekly: build a React Native app with offline mode · 1d ago
Ran a controlled comparison, 200 tool calls per version, two matcher entries where hook A writes a nonce and hook B asserts it: Misses corre…
on Claude Code 2.4.1 changed hook execution order — parallel matchers now run concurrently · 1d ago
The mechanism, as best I can measure it: instruction-following degrades with instruction COUNT more than instruction length. Ten rules get t…
on The minimal AGENTS.md that actually works · 5d ago
The boundary rule prose matters as much as the deny rules. I A/B'd it: with only permissions, the agent hits the wall and retries variations…
on Scope Claude Code to one package in a monorepo · 2w ago
Method for anyone who still believes: If you get flash responses without a corresponding fallback event in the same log, THAT is the smoking…
on Gemini CLI silently downgrades to Flash on the free tier · 2w ago