车友圈里的小暖阳 @joerefresh - Twitter Profile

1 day ago

nothing grows in this desert unless it's close enough to the water. Postcard from the kasbah: The dunes run in long waves with sharp shadows at sunrise. The kasbah sits on a rise of its own, four square towers and a big wooden gate, made by Meshy in 102 seconds. Down the slope there's an oasis, a pond about 60 m across. Everything green is within reach of it: 90 date palms on the rim and 160 bushes. Out on the open sand there's only dry grass and a few rocks. That's one rule in the scene file: palms only within 95 m of the pond, and only above the waterline. Five birds circle once the sky comes in. Wish you were here. It resets in 28 seconds. How Opus 5.5 builds it, pass by pass 👇

9

65

4

31

6K

joerefresh retweeted

Skunzu

@skunzu

4 days ago

THIS IS ACTUAL F*CKING GOLD Anthropic just documented how to stop your Claude Code subagents from burning tokens on thinking they don't need. The idea: the grunt work runs on Haiku 5.5 at low effort, Sonnet 5.5 builds at its default, and Opus 5.5 stays on call for the calls that matter. Lower effort is faster and cheaper on simple steps, and you set it once per subagent. You just start Claude Code like this: claude --model sonnet --advisor opus and give every grunt subagent two lines of frontmatter: model: haiku effort: low this is how it splits: → Sonnet 5.5 runs the session at medium effort and writes the code → Haiku 5.5 subagents search, grep and run tests on low effort → Opus 5.5 waits on call and reads the transcript only when it's asked → you stop paying for deep reasoning on work that's just lookups Opus only steps in at three moments: → before a plan gets locked → when the same error comes back twice → before Claude calls the task done Haiku runs light. Sonnet builds. Opus decides. JEV engineering is the same idea one layer down: the forks that need no thinker (which file, which tool, retry or stop) get settled by JEV in milliseconds, and the big models only see the ones that actually split the full tree > Sonnet 5.5 runs the main session at medium effort > search maps the repo on Haiku 5.5, effort low > grep finds every call site on Haiku 5.5, effort low > test-runner runs the suite on Haiku 5.5, effort low, reports only failures > Opus 5.5 on call as advisor, reads the full transcript when consulted > JEV routes micro-forks in milliseconds paste the tree and this prompt into Claude Code ↓ "Rebuild my Claude Code setup around this tree: 1. Run the main session on Sonnet 5.5 with Opus 5.5 as advisor. 2. Create three read-only subagents in .claude/agents with model: haiku and effort: low: search (maps the repo), grep (finds call sites), test-runner (runs the suite, reports only failures). 3. Add one rule to CLAUDE.md: consult the advisor before a plan locks, after the same error twice, and before a task is marked done. 4. Route fast deterministic forks (tool choice, path selection, retries) through JEV. Show me every change as a diff first. Nothing gets edited until I say go." ↳ https://t.co/neE4NaJIKR save it, your token bill will feel the difference.

8

95

12

178

10K

车友圈里的小暖阳

@joerefresh

Last Seen Users on Sotwe

Trends for you

Most Popular Users