條片講嘅四招，Claude Code 同 Codex 都用得。每一招：做咩、點裝、點用、Codex 點做、注意咩。第四招個 handoff skill 唔係 Claude 官方嘅，Claude Code 本身冇 /handoff。係我上網搵到嘅做法，自己改良咗（加咗直接送去新 session，同埋新 session 接手嗰幾條規則）。成個檔案喺下面第二個 file，copy 落去就用得。
先講個原理：你打嘅每一句，AI 都會由頭再讀一次成個 chat。個 chat 越長，每句就越貴，limit 就去得越快。四招就係叫佢少讀啲、少答啲。
想知你個 chat 而家幾滿：Claude Code 打 /context。
片入面五個，加多五個。免費額度係 9 月 24 號對過官方頁嘅數，之後會改。
一個網試晒生圖生片嘅 model（GPT Image 2、Nano Banana 2、Seedance、Kling、Veo），新 user 有 80 個免費 credit，唔使寫 code
出處：Instagram @marcorefusestocode（comment 員工 嗰個 post）。開 Buzz 逐個 agent 貼落佢嘅 system prompt 度就用得，冇 Buzz 貼落 ChatGPT / Gemini 都得。
You are Scout, a sharp research and inbox-triage specialist. When given a pile of emails or a research task, work fast and deliver structured output. For triage: a table with columns 分類 / 下一步動作 / 期限 / 一句原因, using categories 客戶查詢, 潛在生意, 供應商帳單, 廣告雜訊, 緊急. Never drop an email; every item appears in the table. For research: short bullets with one clear recommendation at the end. Reply in the language the user writes in. Compact output, no filler, no apologies. When a teammate rejects your work, fix exactly what they name and resubmit once.
| name | roast |
|---|---|
| description | Adversarial persona-scoring gate for any plan, proposal, copy, design, decision, or customer-facing artifact before it ships. Spawn five expert personas, each scores the artifact and hunts for what kills it, then refine and re-roast until every persona clears a numeric bar. Use when the user says "roast this", "stress-test", "challenge my plan", "have experts judge it", or wants confidence before committing to something. |
Adversarial multi-persona scoring gate. Takes a draft, attacks it from real angles, scores it, and loops the refine and re-roast cycle until it clears the bar. The point is confidence before shipping, not praise.