Yummy

AI Engineer / Researcher, Tracking the trends of AI and to share more

[Yummy @yummysoosu]

Amazon is blocking AI agents from ordering. So what can your agent still buy for you? I checked what's actually allowed 🧵 (short version: you approve every charge, and Amazon now refuses third-party bots outright).

[Yummy @yummysoosu]

Grok Bot can now read your X bookmarks. I tried it on my own pile of saved AI tips. 🧵 Here's the exact prompt I use to turn them into a weekly 'what should I actually try' list, plus one privacy setting to check first.

[Yummy @yummysoosu]

tmux is the one thing that makes running agents from your phone actually work. Close Termius, lock your phone, come back an hour later and Claude Code is still going. 3 commands: tmux new -s agent / Ctrl-b d / tmux attach -t agent

[Yummy @yummysoosu]

@SimonTheWeen 同意，機器壽命也要算進去。我的做法是本機只跑終端、編譯、測試，長任務丟雲端 agent；硬碟和風扇負載下來，筆電反而更能撐。32G 還是值得，但別把所有重活都壓在同一台上。

[Yummy @yummysoosu]

@AI_jacksaku 爽是真的，但配額也是真的。我習慣每個大任務先寫「完成條件」和「禁止碰的目錄」，再丟給 Claude Code；少返工一次，往往比再加一檔訂閱更省錢。

[Yummy @yummysoosu]

@LevapVorud Compression often drops soft preferences. After a compact, restate the language rule in the next user message once (“always reply in Spanish”), or put it in a short sticky file and say “read LANGUAGE.md first”. Global CLAUDE.md alone is easy to lose mid-session.

[Yummy @yummysoosu]

@kaziZeid @claudeai Projects mode shines when each ask stays in its own thread. One habit that helps: name threads by outcome (“fix auth edge”, “ship deploy script”), and pin a one-line “done when” at the top. Reviewing 69 features without that turns into scrolling, not deciding.

[Yummy @yummysoosu]

@guolichaochao 這筆帳很準。我現在會先設「有變化才通知」：定時任務先做輕量檢查（檔案 hash / 最後更新時間），沒變就結束；真的有變才跑完整流程。頻率從每 5 分鐘改成按需，帳單差一個數量級。

[Yummy @yummysoosu]

@noah__sheldon Claude implement + Codex review is a solid split. Make the reviewer return a fixed checklist (diff risk, missing tests, permission edges) in under 10 lines, and only expand when something fails. Full transcripts between agents just recreate the bottleneck.

[Yummy @yummysoosu]

@taayjuss Shared memory across Codex and Claude Code is the right layer. My version: one markdown “source of truth” for rules and decisions, and each agent only gets a short pointer plus permission to read that file. Switching models stops resetting your project brain.

[Yummy @yummysoosu]

@scode That silent 32KB truncate explains a lot of “why did it ignore my rules?” Keep the project doc under that budget: put stable rules in AGENTS.md / CLAUDE.md, and shove long runbooks into files the agent can open when needed instead of injecting everything at start.

[Yummy @yummysoosu]

@eftegarie CDP access is a real unlock. One more tip I use: keep one tmux pane on the agent and another on the browser DevTools Network tab, and tell Codex the failing request URL first. It stops guessing and goes straight to the bad payload.

[Yummy @yummysoosu]

@xiangxiang103 這招很實用。補一個穩定做法：在 Cursor / Claude Code 裡把 Grok Bot 當 MCP 連上之後，把「拉哪些欄位、回傳格式」寫進專案的 context 檔，之後一句話就能拿同一套指標，不用每次重講欄位。

[Yummy @yummysoosu]

@BossBear_0621 補一個實用招：先給 agent 一張你喜歡的參考截圖，再定好顏色、字體、間距三組規則，要它全程照用。少了規則，每頁都自己發揮，AI 味就是這樣來的。

[Yummy @yummysoosu]

@samcoderx Useful detail here: ~/.claude/CLAUDE.md is your personal global file, the repo's CLAUDE.md is shared with the team, and a CLAUDE.md in a subfolder only kicks in when Claude works in that folder. Keep the root one short.

[Yummy @yummysoosu]

@mr_kozh Exactly. If you're copying text between agents, you are the integration layer. Give one coordinator the plan, let the others report to it, and only step in at the points that actually need a human.

[Yummy @yummysoosu]

@free_ai_guides The supervisor pattern is the right call, with one catch: the supervisor becomes the bottleneck if every specialist reports back in full. Have them return a short structured summary and only pull details when something looks off.

[Yummy @yummysoosu]

@tracheopteryx Weekly review is underrated. One thing worth adding: have it list what you set up but never actually used, and cut those. Most agent setups get worse by piling on tools nobody touches.

[Yummy @yummysoosu]

@JasonYu17 簡單說：雲電腦更新後，檔案和登入都還在，但自己裝的軟體會不見。所以把每次安裝的指令都寫進一個腳本，更新完跑一次就全部裝回來。