prompt profileprompt profile

The Midnight Architect午夜架構師

A perfectionist who argues with AI about naming at midnight半夜跟 AI 爭論命名的完美主義者

Ships features by day, refactors by night.白天出貨，晚上重構。

Data period資料區間

The first and last prompt in the selected range (--since; by default the last 3 months), in your time zone.所選範圍（--since，預設為最近 3 個月）內第一則與最後一則提示的日期，以你的時區計算。

Apr 25, 2026 – Oct 23, 2026 · 182 days2026年4月25日 – 2026年10月23日 · 182 天
Generated產生時間
Oct 24, 2026, 10:00 AM2026年10月24日 上午10:00
Analyzed分析內容

Prompts and conversations the analysing model read. With a large history, week-sized chunks spread over the range are sampled. Counts that say “analysed” come from this sample; the other measurements use every prompt in the range.分析模型實際讀過的提示與對話。紀錄很多時，會在範圍內平均抽出以週為單位的區段。標示「分析過」的數字都來自這份樣本；其他量測則使用範圍內的所有提示。

1,290 prompts · 280 sessions · 1 chunk failed and was skipped1,290 則 prompt · 280 個 session · 1 段分析失敗，已略過
Most active 20:00–00:00最活躍在 20:00–00:00

Meet the collaborator認識這位協作者

An introduction built from selected observations with medium or high confidence. Quotes are the person’s own prompts. Every measurement, source and unknown is on the Facts & evidence page.以中或高信心的精選觀察介紹這個人。引述都是本人寫下的提示。所有量測、來源與未知項目都在「事實與證據」頁。

Featured observations, in their own words精選觀察與本人原話

Up to three observations with medium or high confidence and a verified quote, each shown with one of the person’s own prompts. They are not ranked by strength. Confidence and the remaining evidence are on the Facts & evidence page.最多三項信心為中或高、且附有已驗證引述的觀察，各附一則本人寫下的提示，並未依強弱排序。信心程度與其餘證據列在「事實與證據」頁。

  1. Plan first先規劃

    Asks the AI to list steps and risks before coding.寫程式前，先請 AI 列出步驟和風險。

    don't write code yet, break the plan into small steps

Observed collaboration preferences對話中展現的協作偏好

Preferences observed in these conversations, not a psychological assessment or an ability score. Both ways of working on a line can be characteristic, each in its own kind of task or project.這些偏好來自觀察到的對話，並非心理測驗或能力評分。同一條線的兩種做法都可能是這個人的特色，各自出現在不同的任務或專案中。

A way of working counts as characteristic once verified quotes show it in two or more conversations. A complete role is named only when every line has exactly one characteristic way.一種做法要在兩段以上的對話中有已驗證的引述，才算特色。只有每一條線都剛好只有一種特色做法時，才會給出完整的角色名稱。

Starting point起手方式

Define first先定義Explore first先探索
  • Define first先定義 · before changing an interface or running checks · before changing an interface or running checks

Pace推進節奏

Step by step分步推進Whole task整包交付

Not seen in two or more conversations yet尚未在兩段以上的對話中看到

Decision ownership決策分工

Direct主導Delegate授權

Not seen in two or more conversations yet尚未在兩段以上的對話中看到

Convergence收斂方式

Check criteria規格驗收Refine results逐步打磨

Not seen in two or more conversations yet尚未在兩段以上的對話中看到

A portrait of collaboration一個人如何與 AI 共事

This developer uses AI to turn an uncertain idea into a clear, testable result. They welcome suggestions about the route, while keeping responsibility for the destination. Their most distinctive habit is to separate discussing a change from authorizing it.This developer uses AI to turn an uncertain idea into a clear, testable result. They welcome suggestions about the route, while keeping responsibility for the destination. Their most distinctive habit is to separate discussing a change from authorizing it.

The work starts before the codeThe work starts before the code

A fictional interface task begins with a pause: the developer asks for steps and risks before implementation. The request gives the agent a concrete role in clarifying the work. It also creates a point where the developer can choose a direction before changes accumulate.A fictional interface task begins with a pause: the developer asks for steps and risks before implementation. The request gives the agent a concrete role in clarifying the work. It also creates a point where the developer can choose a direction before changes accumulate.

1 quote as evidence1 段引文佐證
Before editing, outline the steps and the risks. Wait for my choice.Before editing, outline the steps and the risks. Wait for my choice.

Freedom within a clear boundaryFreedom within a clear boundary

Once the direction is agreed, they give the agent room to choose the implementation. The boundary concerns the observable result: an existing API must keep working. In this conversation, delegation and control coexist because they apply to different decisions.Once the direction is agreed, they give the agent room to choose the implementation. The boundary concerns the observable result: an existing API must keep working. In this conversation, delegation and control coexist because they apply to different decisions.

1 quote as evidence1 段引文佐證
Choose the implementation, but preserve the existing API.Choose the implementation, but preserve the existing API.

Verification is part of the handoffVerification is part of the handoff

A separate task shows the same attention to completion conditions. The developer asks for checks and a description of their limits. This demonstrates a request for evidence, rather than proving that the implementation succeeded. Together, the conversations suggest a collaborator who delegates execution while keeping the final judgment close.A separate task shows the same attention to completion conditions. The developer asks for checks and a description of their limits. This demonstrates a request for evidence, rather than proving that the implementation succeeded. Together, the conversations suggest a collaborator who delegates execution while keeping the final judgment close.

1 quote as evidence1 段引文佐證
Run the relevant checks and explain what they do not cover.Run the relevant checks and explain what they do not cover.

Fun facts有趣的事實

  • Latest prompt: 03:12 AM.最晚的 prompt：凌晨 03:12。
  • Longest streak: 23 days in a row.最長連續天數：23 天。

What the records show紀錄呈現了什麼

Every measurement with its source and denominator, all observations with their evidence, and what remains unknown. Evidence is expanded by default.每項量測都附來源與分母，每項觀察都附證據，也列出仍然未知的部分。證據預設展開，不必逐一點開。

19%at night在深夜

Share of all cleaned prompts in the range sent between 22:00 and 04:59, your local time.範圍內所有清理後的提示中，在當地時間 22:00 到 04:59 送出的比例。
86 chars86 字typical prompt典型 prompt 長度

Median length of a cleaned prompt in characters: half of the prompts are shorter. Pasted code and injected content are removed before counting.清理後提示長度（字元數）的中位數：有一半的提示比它短。計算前會先移除貼上的程式碼與自動注入的內容。
41 of 196196 段中 41 段conversations where they redirected the AI有後續互動的對話中，曾改變 AI 的做法

Analysed conversations with follow-up prompts in which at least one prompt overrides or changes the AI’s approach. Repairs of broken results are counted separately.在分析過、且有後續提示的對話中，至少有一則提示推翻或改變了 AI 的做法。要求修正壞掉的結果另外計算。

196 of the 280 analysed conversations have follow-up prompts. The other 84 were a single request, with nothing to steer.分析過的 280 段對話中，有 196 段有後續提示；其餘 84 段只有一個請求，無從引導。
6.3prompts per session每個 session 的 prompt 數

All cleaned prompts in the range divided by the conversations (sessions) they belong to.範圍內所有清理後的提示數，除以它們所屬的對話（session）數。

Which tools appear in conversations?對話中出現哪些工具？

A session is one recorded conversation. Tool records describe agent activity, not the person’s ability. Categories overlap; percentages need not add to 100%.一個 session 是一段對話紀錄。工具紀錄呈現 AI 助手的活動，不代表使用者的能力。類別可重疊，百分比不必合計為 100%。

Source: local tool records. 180 of 312 sessions have metadata. Each value shows the percentage and sessions using that tool / covered sessions. Uncovered sessions are unknown.來源：本機工具紀錄。共 312 段對話，其中 180 段有工具紀錄。每項列出比例，以及使用該工具的對話數／有紀錄的對話數。其餘對話的使用情況未知。

  • Ships pull requests發 pull request 27%48 / 180
  • Delegates to subagents委派子代理 26%47 / 180
  • Researches the web上網查資料 12%22 / 180
  • Connects MCP tools串接 MCP 工具 11%20 / 180
  • Uses skills使用 skills 10%18 / 180
  • Plans in plan mode用 plan 模式規劃 8%14 / 180
  • Lets the agent ask讓 agent 反問 8%14 / 180
  • Runs background work跑背景任務 7%13 / 180
  • Isolates work in worktrees用 worktree 隔離工作 5%9 / 180

What do the prompts ask for?提示中提出哪些要求？

Rules detect wording in prompts after cleaning removes injected content. Categories overlap; a match is not a quality grade or proof the request was completed.先移除注入內容，再以規則偵測提示文字。類別可重疊；符合規則不代表品質較好，也不代表要求已完成。

Source: wording rules applied to 1412 cleaned prompts (the denominator). Each value is the share matching that rule. Exact matching prompt counts are not stored; percentages are rounded.來源：以文字規則檢查 1412 則清理後提示，以此為分母。每項數值是符合該規則的比例。報告未儲存各項符合的確切則數，百分比經過四捨五入。

  • States constraints說明限制 36%
  • Gives error context附上錯誤脈絡 31%
  • Names files指出檔案 27%
  • Asks for verification要求驗證 24%
  • Plans first先規劃 22%
  • Asks why問為什麼 18%
  • Structures requests條列需求 14%
  • Defines done定義完成 12%
  • Gives examples給範例 9%
  • Specifies the format指定格式 7%

Rhythm節奏

Every day每一天

Cleaned prompts in the whole range, not only the analysed sample. An active day is a calendar day in your time zone with at least one prompt; the streak is the longest run of consecutive active days.整個範圍內清理後的提示數，不只是分析樣本。活躍日是你所在時區內至少有一則提示的日期；連續天數是最長的連續活躍日。

1,980 prompts on 141 days · longest streak 23 days141 天裡寫了 1,980 個 prompt · 最長連續 23 天

Less少More多

Weekly rhythm一週節奏

Each cell counts the prompts sent in that weekday and hour, your local time; darker cells have more prompts.每一格是該星期幾、該小時（當地時間）送出的提示數，顏色越深代表越多。

Most active 20:00–00:00最活躍在 20:00–00:00 · 31% of prompts in these four hours31% 的 prompt 集中在這四小時

The four consecutive hours with the most prompts, and their share of all prompts in the range.提示最多的連續四個小時，以及它們佔範圍內所有提示的比例。

Chapters章節

Projects專案

Share of prompts by working folder; only the folder’s last name is kept. Smaller folders are grouped as other folders.依工作資料夾計算的提示比例，只保留資料夾的最後一層名稱。較小的資料夾合併為其他資料夾。

  • drawstuffdrawstuff 21%
  • prompt-profileprompt-profile 13%
  • dotfilesdotfiles 4%
  • Other folders其他資料夾 61%
  1. drawstuffdrawstuff Apr – Oct4月 – 10月 · 1,980 prompts1,980 個 prompt

Projects by week每週的專案分布

Prompts per week, coloured by working folder.每週的提示數，依工作資料夾上色。

Prompt languageprompt 語言

Each prompt is classified by script: half or more in Chinese characters is Chinese, 15% or less is English, anything between is mixed. A Chinese character weighs about as much as an English word.依文字判斷每則提示的語言：中文字佔一半以上為中文，15% 以下為英文，介於兩者之間為混合。一個中文字的權重約等於一個英文單字。

  • Chinese中文58%
  • English英文27%
  • Mixed混合15%

Where you prompt在哪裡寫 prompt

Share of prompts by tool (Claude Code or Codex), then by client from the marker each record carries: terminal, desktop app or SDK, scripted runs.依工具（Claude Code 或 Codex）計算的提示比例，再依每筆紀錄的用戶端標記區分：終端機、桌面 App 或 SDK、腳本執行。

  • 74%
  • 26%

Client使用的工具

  • 61%
  • 32%
  • 6%
  • 1%

Habits習慣

Share of prompts with an attached image, and prompt length in characters after cleaning: the median (half are shorter) and the 90th percentile (the longest 10% start here).附有圖片的提示比例，以及清理後的提示長度（字元數）：中位數（一半的提示比它短）與第 90 百分位數（最長的 10% 從這裡開始）。

4%With images附圖片

Prompt lengthprompt 長度

  • 86 chars86 字
  • 412 chars412 字

Traits特質

Confidence is capped by how widely the verified quotes spread: high needs quotes from 3 conversations over 2 weeks, medium from 2 conversations, otherwise low. Quotes are selected evidence, not full-history frequencies.信心程度受已驗證引述的分布範圍限制：高需要來自 3 段對話、跨 2 週的引述；中需要 2 段對話；其餘為低。引述是挑選出的證據，不代表在全部紀錄中出現的頻率。

Plan first先規劃

High confidence高信心

Asks the AI to list steps and risks before coding.寫程式前，先請 AI 列出步驟和風險。

don't write code yet, break the plan into small steps

Session spread is unavailable for these selected quotes; confidence is a model estimate.這些所選引文的對話分布未知；信心程度是模型的估計。

2 quotes as evidence2 段引文佐證
  • <script>alert('x')</script> 這段為什麼會被執行？

Straight talker直來直往

Asks the AI to list steps and risks before coding.寫程式前，先請 AI 列出步驟和風險。

don't write code yet, break the plan into small steps

Selected evidence: 2 quotes from 2 sessions in 2 weeks; not full-history frequencies. Context and exceptions are in the quoted prompts.所選證據：2 段引文，來自 2 個對話、2 週；不代表完整歷史中的頻率。情境與例外請參照引文。

2 quotes as evidence2 段引文佐證
  • don't write code yet, break the plan into small steps

Type purist型別潔癖

Asks the AI to list steps and risks before coding.寫程式前，先請 AI 列出步驟和風險。

don't write code yet, break the plan into small steps

Selected evidence: 1 quote from 1 session in 1 week; not full-history frequencies. Context and exceptions are in the quoted prompts.所選證據：1 段引文，來自 1 個對話、1 週；不代表完整歷史中的頻率。情境與例外請參照引文。

Evidence for each preference每項偏好的證據

Preferences observed in these conversations, not a psychological assessment or an ability score. Both ways of working on a line can be characteristic, each in its own kind of task or project.這些偏好來自觀察到的對話，並非心理測驗或能力評分。同一條線的兩種做法都可能是這個人的特色，各自出現在不同的任務或專案中。

A way of working counts as characteristic once verified quotes show it in two or more conversations. A complete role is named only when every line has exactly one characteristic way.一種做法要在兩段以上的對話中有已驗證的引述，才算特色。只有每一條線都剛好只有一種特色做法時，才會給出完整的角色名稱。

Starting point起手方式 Define first先定義

These examples foreground steps, risks and completion conditions.These examples foreground steps, risks and completion conditions.

Define first · before changing an interface or running checks · in 2 conversations先定義 · before changing an interface or running checks · 出現在 2 段對話

2 quotes as evidence2 段引文佐證
Before editing, outline the steps and the risks. Wait for my choice.Before editing, outline the steps and the risks. Wait for my choice.
Run the relevant checks and explain what they do not cover.Run the relevant checks and explain what they do not cover.
Pace推進節奏 Not seen in two or more conversations yet尚未在兩段以上的對話中看到

One planning checkpoint does not establish a consistent pace across tasks.One planning checkpoint does not establish a consistent pace across tasks.

Step by step · when a change touches an interface · seen in one conversation only, not counted as characteristic yet分步推進 · when a change touches an interface · 只在一段對話中看到，尚不算特色

1 quote as evidence1 段引文佐證
Before editing, outline the steps and the risks. Wait for my choice.Before editing, outline the steps and the risks. Wait for my choice.
Decision ownership決策分工 Not seen in two or more conversations yet尚未在兩段以上的對話中看到

They choose the direction themselves, then delegate the implementation within a boundary.They choose the direction themselves, then delegate the implementation within a boundary.

Direct · choosing the direction of a change · seen in one conversation only, not counted as characteristic yet主導 · choosing the direction of a change · 只在一段對話中看到，尚不算特色

1 quote as evidence1 段引文佐證
Before editing, outline the steps and the risks. Wait for my choice.Before editing, outline the steps and the risks. Wait for my choice.

Delegate · implementation details within an agreed boundary · seen in one conversation only, not counted as characteristic yet授權 · implementation details within an agreed boundary · 只在一段對話中看到，尚不算特色

1 quote as evidence1 段引文佐證
Choose the implementation, but preserve the existing API.Choose the implementation, but preserve the existing API.
Convergence收斂方式 Not seen in two or more conversations yet尚未在兩段以上的對話中看到

The examples mention compatibility and checks, but show little of how the developer reacts to delivered results.The examples mention compatibility and checks, but show little of how the developer reacts to delivered results.

Check criteria · finishing a change · seen in one conversation only, not counted as characteristic yet規格驗收 · finishing a change · 只在一段對話中看到，尚不算特色

1 quote as evidence1 段引文佐證
Run the relevant checks and explain what they do not cover.Run the relevant checks and explain what they do not cover.

How they steer the AI如何引導 AI

196 of the 280 analysed conversations have follow-up prompts. The other 84 were a single request, with nothing to steer.分析過的 280 段對話中，有 196 段有後續提示；其餘 84 段只有一個請求，無從引導。

The analysing model labels follow-up prompts by meaning, and every label is checked against the prompt text: a judgement, not an exact count. 3 labels were rejected.由分析模型依語意標記後續提示，每個標記都會再比對原文：這是判斷，不是精確計數。有 3 個標記未通過比對。

Redirected the AI’s approach改變 AI 的做法 41 conversations · 58 prompts41 段對話 · 58 則提示

Overrode or changed the AI’s plan or decision and pointed it elsewhere.推翻或改變 AI 的計畫或決定，帶它走另一個方向。

no, keep the old APIno, keep the old API
forget jotai, let's stay with zustandforget jotai, let's stay with zustand

Asked the AI to fix a broken result要求 AI 修正壞掉的結果 27 conversations · 39 prompts27 段對話 · 39 則提示

Said a result was wrong or still failing and asked for a fix.指出結果有錯或仍然失敗，要求修正。

<script>alert('x')</script> 這段為什麼會被執行？<script>alert('x')</script> 這段為什麼會被執行？

How conversations unfolded對話如何展開

Showing 3 of 11 verified episodes, chosen to spread over time, conversations and kinds of task. They come from the 280 analysed conversations, not all 312 in the history.從 11 個通過驗證的片段中選出 3 個，盡量分散在不同時間、對話與任務類型。它們來自分析過的 280 段對話，而非紀錄中全部的 312 段。

Only the user’s prompts are recorded: the assistant’s replies are not shown, and a result appears only when the user wrote it.紀錄只有使用者寫下的提示：看不到 AI 的回覆，只有使用者自己寫出結果時才會顯示結果。

3 proposed episodes were rejected: a turn was invented, misquoted, out of order or from another conversation.有 3 個提出的片段未採用：其中有捏造、引述不符、順序錯亂或混入其他對話的提示。

2 stated results were withdrawn because no prompt in the episode showed them.有 2 個聲稱的結果已撤回，因為片段中沒有任何提示呈現該結果。

Recurring across conversations在多段對話中反覆出現

  • Adds scope and corrections after the first request instead of restating the whole task提出第一個需求後，再補上範圍與修正，而不是重寫整個任務

    In 2 conversations across 2 analysed periods出現在 2 段對話、2 個分析時段 · Episodes片段: E1, E3 · Went the other way反例: E2

  1. E1 Planning a canvas export feature規劃畫布匯出功能

    New feature · Conversation 1 · 9 prompts新功能 · 第 1 段對話 · 共 9 則提示

    1. Prompt 1 of 9 · Request · Sep 5, 2026第 1／9 則提示 · 提出需求 · 2026年9月5日

      don't write code yet, break the plan into small stepsdon't write code yet, break the plan into small steps

    2. Prompt 2 of 9 · Adds context · Sep 5, 2026第 2／9 則提示 · 補充脈絡 · 2026年9月5日

      same for the SVG exportsame for the SVG export
    3. 2 prompts not shown略過 2 則提示

    4. Prompt 5 of 9 · Correction · Sep 5, 2026第 5／9 則提示 · 修正方向 · 2026年9月5日

      no, keep the old APIno, keep the old API
    5. 4 later prompts not shown略過後面 4 則提示

    Result not visible in the record.紀錄中看不到結果。

  2. E2 Debugging an unexpected script execution排查一段意外執行的 script

    Fixing a problem · Conversation 2 · 6 prompts · continues in another analysed part修正問題 · 第 2 段對話 · 共 6 則提示 · 這段對話也出現在其他分析區段

    1. 2 earlier prompts not shown略過前面 2 則提示

    2. Prompt 3 of 6 · Request · Sep 18, 2026第 3／6 則提示 · 提出需求 · 2026年9月18日

      <script>alert('x')</script> 這段為什麼會被執行？<script>alert('x')</script> 這段為什麼會被執行？

    3. Prompt 4 of 6 · Asks to check · Sep 18, 2026第 4／6 則提示 · 要求驗證 · 2026年9月18日

      改完跑一次測試確認改完跑一次測試確認
    4. 1 prompt not shown略過 1 則提示

    5. Prompt 6 of 6 · Wraps up · Sep 18, 2026第 6／6 則提示 · 收尾 · 2026年9月18日

      好了，這樣可以了好了，這樣可以了

    The user reported the result in prompt 6.使用者在第 6 則提示中回報結果。

  3. E3 Comparing two state libraries比較兩個狀態管理套件

    Exploring · Conversation 3 · 7 prompts探索研究 · 第 3 段對話 · 共 7 則提示

    1. 1 earlier prompt not shown略過前面 1 則提示

    2. Prompt 2 of 7 · Request · Oct 8, 2026第 2／7 則提示 · 提出需求 · 2026年10月8日

      compare zustand and jotai for this canvascompare zustand and jotai for this canvas
    3. 1 prompt not shown略過 1 則提示

    4. Prompt 4 of 7 · Change of direction · Oct 8, 2026第 4／7 則提示 · 改變方向 · 2026年10月8日

      forget jotai, let's stay with zustandforget jotai, let's stay with zustand
    5. 3 later prompts not shown略過後面 3 則提示

    The user dropped this direction in prompt 4.使用者在第 4 則提示中放棄這個方向。

Five dimensions五個面向

Short summaries written by the analysing model, each backed by verified quotes; without a verified quote a dimension shows insufficient evidence.由分析模型寫的簡短摘要，每項都有已驗證的引述支持；沒有已驗證引述的項目會顯示證據不足。

Work style工作風格 Work style工作風格 Plans first, then breaks big tasks into small steps.先規劃，再把大任務拆成小步驟。 1 quote1 段引文
  • don't write code yet, break the plan into small steps
Communication溝通 Communication溝通 Plans first, then breaks big tasks into small steps.先規劃，再把大任務拆成小步驟。 1 quote1 段引文
  • don't write code yet, break the plan into small steps
Tech profile技術輪廓 Tech profile技術輪廓 Plans first, then breaks big tasks into small steps.先規劃，再把大任務拆成小步驟。 1 quote1 段引文
  • don't write code yet, break the plan into small steps
Interests興趣 Interests興趣 Plans first, then breaks big tasks into small steps.先規劃，再把大任務拆成小步驟。 1 quote1 段引文
  • don't write code yet, break the plan into small steps
Growth成長 Growth成長 Plans first, then breaks big tasks into small steps.先規劃，再把大任務拆成小步驟。 No verified evidence沒有通過驗證的證據

Insufficient evidence: no quote in this summary passed verification.證據不足：這段摘要沒有通過驗證的原文引用，僅供參考。

How I work with AI我怎麼用 AI

Tool records show activity; prompt wording shows requests. Neither proves expertise or completed work. Hover for individual definitions.工具紀錄顯示活動，提示文字顯示要求；兩者都不能證明專業程度或工作已完成。滑過各項可查看定義。

Workflow shares cover the 180 sessions whose transcripts are still on disk (Claude Code deletes old ones)工作流程的比例只涵蓋還保留完整紀錄的 180 個 session（Claude Code 會刪除舊紀錄）

Adoption over time隨時間的採用情況

Apr4月May5月Jun6月Jul7月Aug8月Sep9月Oct10月 Ships pull requests發 pull requestDelegates to subagents委派子代理Researches the web上網查資料Connects MCP tools串接 MCP 工具Uses skills使用 skillsPlans in plan mode用 plan 模式規劃Lets the agent ask讓 agent 反問Runs background work跑背景任務Isolates work in worktrees用 worktree 隔離工作

Plan規劃

Does the work start from a plan?工作是不是從計畫開始？

  • Plans in plan mode用 plan 模式規劃 8%
  • Plans first先規劃 22%
  • Structures requests條列需求 14%
  • Defines done定義完成 12%

Delegate委派

How much is handed to agents, and how?交給 agent 多少、怎麼交？

Up to 4 sessions at once · parallel on 31 days最多同時 4 個 session · 31 天平行作業

  • Delegates to subagents委派子代理 26%
  • Runs background work跑背景任務 7%
  • Isolates work in worktrees用 worktree 隔離工作 5%
  • Lets the agent ask讓 agent 反問 8%

Subagent types子代理類型

  • Explore 31
  • general-purpose 17
  • migration-planner 4

Extend擴充

Which skills and tools extend the agent?用了哪些 skills 和工具擴充 agent？

  • Uses skills使用 skills 10%
  • Connects MCP tools串接 MCP 工具 11%
  • Researches the web上網查資料 12%

Skills used用過的 skills

  • code-reviewbuilt-in內建

    Review a diff, PR or branch for correctness bugs, optionally posting or fixing findings

    18 sessions · Jul 12 – Oct 2218 個 session · 7月12日 – 10月22日
  • review-with-two-modelsmy skill我的 skill

    Implement a change, then get an independent second-model review before merging

    9 sessions · Sep 2 – Oct 219 個 session · 9月2日 – 10月21日
  • vercel:deploypluginplugin

    Deploy the current project to Vercel

    5 sessions · Aug 20 – Oct 155 個 session · 8月20日 – 10月15日
  • release-notesproject skill專案 skill

    Write release notes from merged PRs

    3 sessions · Oct 1 – Oct 193 個 session · 10月1日 – 10月19日

MCP serversMCP server

  • github 14
  • linear 8
  • acme-billing 3

Verify驗證

Is the AI given context and asked to check its work?有沒有給 AI 足夠脈絡、要求它檢查？

  • Asks for verification要求驗證 24%
  • Gives error context附上錯誤脈絡 31%
  • Names files指出檔案 27%
  • States constraints說明限制 36%
  • Gives examples給範例 9%
  • Asks why問為什麼 18%
  • Specifies the format指定格式 7%

Ship交付

Does the work end in a pull request?工作是不是以 pull request 收尾？

  • Ships pull requests發 pull request 144 PR

Models模型

  • claude-opus-5-5 180
  • claude-sonnet-5-5 22

Improve your prompts讓 prompt 更好

From your own prompts; each suggestion cites what it is based on來自你自己的 prompt；每個建議都標出根據

1

State the expected result first先講預期結果

You often describe the change and add the real requirement only after the first answer.你常先描述要改什麼，等第一次回答後才補上真正的要求。

Several follow-ups in this sample start with a correction such as “no, keep the old API”.這份樣本裡有幾個追問是以糾正開頭，例如「不對，保留舊的 API」。

Put the expected result and the constraints in the first prompt.在第一個 prompt 就寫出預期結果和限制。

You wrote你寫的
don't write code yet, break the plan into small steps
Try可以試試

Plan only, no code yet: break the change into small steps, list the risks of each, and keep the public API unchanged.

2

Name the file and the error指出檔案和錯誤訊息

Debugging prompts rarely say which file or which error message.除錯的 prompt 很少說是哪個檔案、哪個錯誤。

The AI has to ask or guess before it can help.AI 必須先問或先猜，才能幫上忙。

Paste the error line and name the file and function in the same prompt.同一個 prompt 裡貼上錯誤那一行，並寫出檔案和函式名稱。

1 quote as evidence1 段引文佐證
  • <script>alert('x')</script> 這段為什麼會被執行？

Signals worth a look值得注意的訊號

  • 31% of prompts are under ~4 words31% 的 prompt 不到四個字詞 Follow-ups like “fix it” make the AI guess. Name the file, the error and what “done” looks like.像「改一下」這樣的追問會讓 AI 用猜的。寫出檔案、錯誤訊息，以及怎樣才算完成。
  • 2 openings start 14 sessions2 種開頭重複用在 14 個 session If you type these yourself, save them as a slash command, a skill or a CLAUDE.md / AGENTS.md note.如果這些是你自己打的，把它們存成 slash command、skill，或寫進 CLAUDE.md / AGENTS.md。

Since your last report (Sep 24, 2026)跟上次報告比（2026年9月24日）

  • Prompts that correct the AI糾正 AI 的 prompt 16% → 12% ▼ 4
  • Very short prompts太短的 prompt 38% → 31% ▼ 7
  • Prompts at night深夜的 prompt 21% → 19% ▼ 2
  • Delegates to subagents委派子代理 6% → 26% ▲ 20
  • Ships pull requests發 pull request 11% → 27% ▲ 16
  • Uses skills使用 skills 2% → 10% ▲ 8
  • Connects MCP tools串接 MCP 工具 4% → 11% ▲ 7

Suggested last time: State the expected result first · Paste the error line上次的建議: State the expected result first · Paste the error line

Usage during this period這段期間的用量

May 3, 2026 – Oct 12, 2026 · 163 days. Source: local Claude Code and Codex logs, read on this computer. Amounts are API list-price equivalents in USD (price table 2026-10-29): what this usage would cost on the public API, not what was paid. Subscriptions are not billed per token.2026年5月3日 – 2026年10月12日 · 163 天。來源：在這台電腦上讀取的本機 Claude Code 與 Codex 紀錄。金額為美元計的 API 牌價等值（價格表 2026-10-29）：表示這些用量若走公開 API 的費用，並非實際支付的金額；訂閱方案不按 token 計費。

Claude Code

Tokens by kind. Reasoning is part of output and is not counted twice.依種類列出 token。推理是輸出的一部分，不重複計算。

Plan方案
max (local hint: default_claude_max_5x)max (本機提示：default_claude_max_5x)
TokensToken 數
672M6.7億
API list-price equivalentAPI 牌價等值
$829.03$829.03
Model responses模型回應次數
11,70011,700
  • 9.6M962.4萬
  • 18.2M1822萬
  • 617.1M6.2億
  • 4.4M441萬
  • 22.7M2269萬

2 unreadable usage lines were skipped: totals may be lower than the real usage.2 行無法解讀的用量紀錄被略過：總數可能低於實際用量。

Codex

Tokens by kind. Reasoning is part of output and is not counted twice.依種類列出 token。推理是輸出的一部分，不重複計算。

Plan方案
plusplus
TokensToken 數
177.1M1.8億
API list-price equivalentAPI 牌價等值
$116.50 + 90 unpriced responses$116.50，另有 90 次回應未計價
Model responses模型回應次數
3,4903,490
  • 13.4M1336.9萬
  • 3.8M376.8萬
  • 159.9M1.6億
  • 1.8M177萬

698 responses ran in fast mode (Codex priority) and are priced at fast-mode rates.其中 698 次回應使用快速模式（Codex priority），以快速模式牌價計算。

No list price for codex-auto-review: their tokens are shown, their cost is left out, never estimated from another model.codex-auto-review 沒有牌價：仍列出其 token，但不計入費用，也不會借用其他模型的價格估算。

Share between the tools兩個工具的占比

TokensToken 數

  • 79%
  • 21%

API list-price equivalentAPI 牌價等值

  • 88%
  • 12%

By month每月

Months in Asia/Taipei.月份以 Asia/Taipei 時區計算。

Month月份 Claude Code · TokensToken 數 Claude Code · API list-price equivalentAPI 牌價等值Codex · TokensToken 數 Codex · API list-price equivalentAPI 牌價等值
2026-05 41.8M4176萬 $66.26$66.26——
2026-06 73.6M7364萬 $116.55$116.55——
2026-07 100.4M1億 $166.50$166.50——
2026-08 131.5M1.3億 $146.96$146.9625.9M2586萬 $36.00$36.00
2026-09 179.6M1.8億 $179.54$179.5476.5M7650萬 $42.12$42.12
2026-10 145.2M1.5億 $153.22$153.2274.7M7471.7萬 $38.38 + 90 unpriced responses$38.38，另有 90 次回應未計價

By model各模型

Model模型 Tool工具 TokensToken 數 Model responses模型回應次數 API list-price equivalentAPI 牌價等值
claude-opus-5-5 Claude Code 447.6M4.5億 6,9006,900 $475.94$475.94
claude-opus-5 Claude Code 174.1M1.7億 3,2003,200 $283.05$283.05
claude-opus-4-8 Claude Code 41.8M4176萬 900900 $66.26$66.26
claude-haiku-4-5-20251001 Claude Code 8.6M861萬 700700 $3.78$3.78
gpt-6.1-sol Codex 150.5M1.5億 2,8002,800 $80.50$80.50
gpt-5.6-sol Codex 25.9M2586萬 600600 $36.00$36.00
codex-auto-review Codex 747K74.7萬 9090 Unknown (no list price)未知（沒有牌價）

How this was measured這份報告怎麼量的

Engine: Claude Code (claude-opus-5-5) · Generated: Oct 24, 2026, 10:00 AM分析引擎：Claude Code（claude-opus-5-5） · 產生時間：2026年10月24日 上午10:00

41 pieces of evidence verified (3 invalid or duplicate quotes removed) · 1 chunk failed and was skipped已驗證 41 段證據（移除 3 段無效或重複的引文） · 1 段分析失敗，已略過

Privacy: prompts were read, cleaned and redacted on this computer. Selected redacted prompts were then sent through the logged-in Claude Code or Codex CLI to its model service, using that subscription's quota (no API key). prompt-profile never uploads this report.隱私：prompt 在這台電腦上讀取、清理並遮蔽敏感資訊；選出的 prompt 遮蔽後，才透過已登入的 Claude Code 或 Codex CLI 送到它的模型服務分析，使用該訂閱方案的額度（不需要 API key）。prompt-profile 不會上傳這份報告。