WHAT NEW IN X11.05 THU

X 上的 AI 合集档案 · 23 个合集 · 69,408 条

jevVIDEO

8 个免费 GitHub 仓库修复 Claude Code 半成品问题

视频盘点让 agent 真正测试应用、少说废话、按计划验收的免费工具,其中 reticle 被作者称为「Jev 式感知」——返回通过/失败/无法判断的定型裁决而非感觉。

sonicdr1p
@sonicdr1p2026.09.25收录原推发布于 2026.09.25
sonicdr1p @sonicdr1p原推发布于 2026.09.25
在 X 上打开原推

claude code says "done" way more often than your app actually works. this video shows 8 free github repos that fix that, plus a few that make your agent cheaper to run if you use claude code or codex and keep getting half broken apps, this is the upgrade layer almost nobody installs. every repo is free, you only need the agent itself 00:36 turn a photo into a 3D object for your landing page 01:37 make the agent actually test your app before it says done 02:55 shorter replies and less useless code, fewer tokens burned 04:44 a design skills library the agent picks from on its own 06:52 plan the app first, then check the result against the plan 08:22 iphone skills for liquid glass and the new foldable 09:34 find out if your skills help or just eat tokens 11:37 a spellcheck for the code your agent writes my take: skip the 3D one unless you have tokens to spare, the video itself says it burns a lot and takes a while. the 3 i'd steal today: reticle opens your real running app, clicks through it and hands back pass, fail or couldn't tell with the exact file and line to fix. fun detail, the devs call it Jev style perception, same idea as the article i'm quoting: typed verdicts instead of vibes chisle makes the agent talk less and write less code. in the author's own test one task dropped from about 1,500 tokens to about 600 caliper runs your agent with and without a skill and shows if it's worth the tokens. most people never test their skills, they just keep piling them up which one are you installing first

这条是引用着另一条写的:x.com/sonicdr1p/status/2102824068357435496

x.com/sonicdr1p/status/2103166549267759582采集于 2026.09.25 10:05