Make Claude Code give you answers, not essays
YouTube 影片重點整理
Make Claude Code give you answers, not essays
這支 Shorts 示範如何讓 Claude Code 或其他 AI agent 的回覆更容易閱讀：在全域 CLAUDE.md 加上一句要求它極度精簡，甚至為了精簡可以犧牲語法完整度。作者用同一個 debug 任務做對照：沒有這句指令時，Claude 仍會輸出可讀但偏段落化的說明；加入指令後，回覆變成更直接的短標題、短清單與重點摘要。作者強調，模型底層仍理解它做了什麼；只是第一眼收到的回報更快、更好掃讀。
01
把「極度精簡」寫進全域 CLAUDE.md
0:06作者的核心做法很小：在提示詞加上 Be extremely concise. Sacrifice grammar for the sake of concision.
他不是每次臨時輸入，而是把這段要求放進 Claude Code 的全域 `claude.md`，讓後續回覆預設就比較短、比較容易掃讀。
Prompt convention02
用同一個 debug 任務比較差異
0:26作者找了一個需要 Claude debug 的錯誤，先用沒有精簡指令的版本跑一次。結果雖然仍算短，但輸出偏段落式，閱讀上比較像一小段文章。
加入精簡指令後，Claude 直接給出更短的診斷與標題式結論，例如指出 ESM/CJS mismatch、檢查 build output 與 config，並用 `Root cause found` 這類短標題切段。
A/B test03
第一份回報越短，後續追問越快
0:54影片中最大的差異不是 Claude 有沒有理解任務，而是最後摘要的呈現方式：未加指令時像一小篇文章；精簡版則是短句、標題與清單。
作者明確說，如果需要後續問題仍然可以追問；他偏好的是第一份回報先用更少字把資訊送到眼前。
Agent UX回報資訊時：極度精簡；為了精簡，可以犧牲語法。
重點時間戳索引
- 0:00作者說明目標：讓 Claude Code 或任何 agent 的輸出更易讀、更好處理。
- 0:06他最近常在 prompt 加一句：極度精簡，為了精簡可以犧牲語法。
- 0:18他把這句話放進全域 claude.md，讓 Claude 的回饋預設變得更精簡。
- 0:26測試方式是找一個需要 Claude debug 的錯誤，先不加精簡指令執行一次。
- 0:33未加指令的版本會輸出段落式說明；雖然不長，但不夠乾淨。
- 0:41精簡版本直接指出 ESM/CJS mismatch，並用短標題呈現 root cause。
- 0:54最大差異在最終摘要：未加指令像一小篇文章，精簡版則是短句、標題與清單。
- 1:11如果需要追問仍可追問；作者只是偏好第一份回報先保持高度精簡。
- 1:22他的 CLAUDE.md 寫法是：回報資訊時極度精簡，為了精簡犧牲語法。
⚠️ 未確認 / 無法核對項目
- 逐字稿自動字幕將 Claude Code 誤轉為 clawed code / claw. MD；報告主文已依影片標題與上下文更正為 Claude Code / CLAUDE.md。
- 0:41 的 transcript 寫作 ESM CGS mismatch，依常見技術語境判讀為 ESM/CJS mismatch，但原字幕未提供畫面文字核對。
📎 原始內容（查證 / AI 追查用，非閱讀主文，點擊展開）
YouTube description
Become an AI Hero: https://www.aihero.dev Follow Matt on Twitter https://twitter.com/mattpocockuk
YouTube transcript with timestamps
0:00 Here's how you can make your clawed code 0:01 output or really any agent output a lot 0:03 more readable and easier to work with. 0:06 Recently, I've been talking about this 0:07 extra little bit that I add to prompts 0:09 saying, "Be extremely concise. Sacrifice 0:12 grammar for the sake of concision. This 0:14 turns large prompts into really nice 0:16 human readable format." And so I thought 0:18 to myself, why don't I add this into my 0:20 Claude code global claude.md file? 0:23 Because then I'll get much more concise 0:24 feedback coming from Claude. So I 0:26 decided to do a test. I found an error 0:28 that I needed Claude to debug for me and 0:30 I ran it without the extra command to 0:32 sacrifice its grammar. We can see that 0:33 straight away it starts outputting kind 0:35 of paragraphs for me. This is still 0:36 pretty short and easy to scan, but it's 0:39 still not as clean as it could be. I 0:41 think in the concise version, I get ESM 0:43 CGS mismatch. Let me check the build 0:44 output and config. Very direct. It 0:46 starts using nice little headers, too. 0:48 Root cause found. React 19 0:50 incompatibility with ink 4.1.0. Here's 0:52 another one. Still same error, but the 0:54 biggest difference is in the final 0:56 summary. This is without the concise 0:58 instruction where it gives me a fair 1:00 chunk of stuff, a little mini article to 1:02 read here. Whereas in the concise 1:04 version, it really is just a few words 1:06 here. A couple of headings, a little 1:08 list, and we get the information faster. 1:11 Now, of course, if I need to follow this 1:12 up with extra questions, I can. Under 1:15 the hood, the LLM still has the same 1:16 understanding of what it's done, but I 1:18 prefer this concise summary as the first 1:21 thing I receive. And here's what I have 1:22 in my claw. MD file. When reporting 1:24 information to me, be extremely concise 1:26 and sacrifice grammar for the sake of 1:28 concision.