MumbleKeyBlog

← Nobody performs for a coding agent, so I let one read six months of my logs

繁體中文

我讓 AI 挖了我半年的對話紀錄，然後被它當面點破

學長丟來一個他在推特上撿到的 prompt，開頭三句話是這樣：

People lie in journals. They perform in therapy. Nobody performs for a coding agent.

人會在日記裡說謊，會在心理師面前表演，但沒有人會對 coding agent 演戲。叫 agent 改東西的時候不會顧形象，不會演，急了連標點符號都懶得打。所以那份 prompt 主張：你機器上的 agent 對話紀錄，是你這個人最誠實的紀錄。它設計了六個階段，把紀錄挖出來，最後告訴你「你是誰」。

聽起來很玄，玩玩看。

先貼數字

我讓 Claude 跑在我自己的紀錄上：半年，7580 句指令。挖出來的幾個數字：

  • 「驗」這個字出現 438 次，是我打過最高頻的動詞
  • 收工儀式 bye 打了 50 次，handoff 打了 49 次
  • 「卡住了嗎？」問了 10 次
  • 「字幕結尾不要全型句號」同一句修正我講了 5 次
  • /usage 額度查了 783 次

有些數字看了會笑，有些看了會沉默。同一句修正講 5 次，我的教訓管理顯然有洞。額度查了 783 次，我隨時在擔心機器不夠用，卻從來沒算過我自己夠不夠用。

然後它開始問問題

一次一題，不給結論，答案跟紀錄矛盾就把收據拿出來。最後那面鏡子裡有一句話，我看完愣了很久：

你蓋了一座只進不出的城，然後把「出去」這件事，交給了運氣。

它的依據是：我幫 GitHub 趨勢、市場數據、排程健康全都裝了自動雷達，全部指向內部，餵養我自己的系統。指向外面的感測器，一個都沒有。這半年我的機會全靠偶然。

這不是雞湯，每個字後面都有我自己的時間戳。那份 prompt 開頭就把話講完了：建議可以裝沒看到，自己的紀錄沒辦法反駁。

當天做成 skill，開源

被打完之後，當天就把這份 prompt 做成了 skill：

https://github.com/fredchu/session-mirror

做的過程也順便誠實一下：設計是我出嘴，Claude 主持流程，實作和四輪 review 是 Codex 跑的，循環到只剩小問題才放行。統計的部分抽成純 Python 腳本（零依賴），同樣的輸入就該有同樣的數字，這種事不該讓語言模型每次重新發揮。判斷和問問題留給模型，但關卡寫死：它得先交出採礦計畫，你點頭之前一個字都不能讀；所有輸出先過機密遮蔽；要動你的設定檔，逐條 diff 給你批准。

隱私也講清楚：腳本保證只在本機跑、不連網。但 agent 讀進去的內容會進到模型的 context，用雲端模型就等於那些內容會到模型供應商那邊。這件事寫死在第一關的告知裡，想全程本機就用本地模型跑。

致謝

這個 skill 的六階段概念來自 Machina（@EXM7777）2026-07-07 發在 X 的〈This prompt will change your life〉。文章剛發的時候我還找不到人，發完沒多久就找到原推了，署名馬上補上。題頭那三句話是全 repo 唯一逐字引用。謝謝你寫下它。

裝好之後對你的 agent 說「照鏡子」就會開跑。

至於我自己，鏡子照完拿到一份 30 天的路線圖，第一項就是幫「會僱用我的世界」裝一個雷達。城牆上總得開幾扇窗。

English

Nobody performs for a coding agent, so I let one read six months of my logs

A friend sent me a prompt he found on Twitter. It opens like this:

People lie in journals. They perform in therapy. Nobody performs for a coding agent.

Which is true. When I’m telling an agent to fix something at 11pm I am not curating an image. Nobody types “please” on the third retry. So the prompt’s claim: the session logs sitting on your machine are the closest thing to an honest autobiography you will ever produce. It lays out six phases to excavate that record and tell you who you actually are.

Sounded like fortune-cookie territory. I ran it anyway.

The numbers first

I pointed Claude at my own history: six months, 7580 prompts. Some numbers it dug up:

  • “verify” (驗) appears 438 times, my single most-typed verb
  • “bye” typed 50 times, “handoff” 49 times, my end-of-session ritual
  • “are you stuck?” asked 10 times
  • the same subtitle formatting correction, repeated 5 times
  • /usage checked 783 times

Some of these are funny. Some of them shut me up. Repeating the same correction 5 times means my lessons-learned system has a hole in it. Checking quota 783 times means I constantly worry whether the machine has enough left, and never once measured whether I do.

Then it started asking questions

One at a time, no conclusions, and when my answer contradicted the record it produced the receipt. At the end it held up a mirror, and one line in it stopped me cold:

You built a fortress that only takes things in, and left “getting out” to luck.

Its evidence: I have automated radars for GitHub trends, market data, cron health. All of them point inward, feeding my own system. Sensors pointing outward, at readers, at users, at anyone who might hire me: zero. Every opportunity I got in six months arrived by accident.

That’s not a horoscope. Every word carries a timestamp from my own logs. The prompt said it upfront: advice you can ignore, but a true pattern with your own timestamps on it, you cannot.

Same day, skillified and open-sourced

Still a bit stunned, I turned the prompt into a skill that same day:

https://github.com/fredchu/session-mirror

Honest credits on how it was built: I supplied opinions, Claude ran the show, Codex did the implementation plus four review rounds, looping until only nitpicks remained. The counting lives in plain Python scripts (zero dependencies), because the same input should produce the same numbers and that is not a job for a language model’s creativity. The judgment stays with the model, but the gates are hard-coded: it must show you its mining plan and read nothing until you approve; every output passes secret redaction; config changes ship as diffs you approve one by one.

Privacy, stated bluntly: the scripts are local-only, no network, guaranteed. The model is not. Whatever the agent reads enters model context, and on a hosted model that content reaches your provider. This disclosure is hard-coded into the first gate. Run a local model if you want the whole thing on-machine.

Attribution

The six-phase concept comes from “This prompt will change your life” by Machina (@EXM7777), published on X on 2026-07-07. When this post first went up I couldn’t find the source; it turned up shortly after publishing, and the credit went in immediately. The three sentences at the top are the only verbatim quote in the repo. Thank you for writing it.

Once installed, tell your agent “mirror me” and it runs.

As for me, the mirror came with a 30-day roadmap. Item one: build a radar for the world that might hire me. A fortress could use a few windows.