claude-idea-review-skill
Claude Code skill · idea reviewClaude Code skill · 點子驗證

Pressure-test the idea before you build it.

先挑戰點子，
再決定做不做。

Ask Claude Code to validate an idea. Four thinking personalities argue it over 5, 12, or 16 rounds, and Codex, Gemini, or Grok add their own analysis when their CLIs answer. You get a verdict with a kill criterion.

請 Claude Code 驗證一個點子，四種思考個性會辯論 5、12 或 16 輪；Codex、Gemini、Grok 的 CLI 能回應時，也會加入各自的分析。最後你拿到的是附停止條件的結論。

Install in one command一行指令安裝View source查看原始碼
MITSKILL.md + Python 3.8+Made for the 自由工坊 community為自由工坊社群而做
Startup check啟動檢查5 rounds5 輪
check_providers.py --smoke
3 optional providers · 45 s per smoke test3 個選配 provider · 每個測試 45 秒逾時
✓Codex answersCodex 有回應codex exec · smoke_ok: trueCLI
!Gemini found, no answer找到 Gemini，但沒有回應agy -p · 測不到帳號key?
…Grok not installed沒有安裝 Grokgrok cli not found · Claude 模擬 GrokClaude
✓Web research saved網路研究已存檔input.md · Research Summaryweb
Asks once for keys, never blockskey 只問一次，不會卡住Round 1 of 5第 1 輪，共 5 輪
5 / 12 / 16Round options, 5 by default可選輪數，預設 5 輪
4Thinking personalities per round每輪的思考個性
3Optional providers it checks會檢查的選配 provider
9Fields in the final verdict最終結論的欄位
How a review runs驗證流程

Research, argue, then decide.

先研究，再辯論，最後下決定。

01

Check providers

檢查 provider

check_providers.py looks for codex, agy or gemini, and grok, and sends each one a one-line prompt. If a CLI is installed but cannot answer, Claude reports it as 測不到帳號 (account not reachable).

check_providers.py 會找 codex、agy 或 gemini，以及 grok，並各送一句測試提示。裝了卻無法回應的 CLI，Claude 會回報「測不到帳號」。

02

Research first

先做研究

With web tools available, Claude gathers competitors, pricing, demand, platform limits, and policy constraints. Later rounds must separate evidence from assumptions.

有網路工具時，Claude 會先蒐集競品、價格、需求、平台限制與政策限制。之後每一輪都要分清楚哪些是證據、哪些是假設。

03

Argue in rounds

分輪辯論

Round 1 is first positions, the middle rounds clash, and the last round converges. After round 1, every personality must answer an earlier point.

第 1 輪各自表態，中間幾輪互相攻防，最後一輪收斂。第 1 輪之後，每個個性都要回應先前的論點。

04

Decide and print

下結論、印報告

final.md holds the verdict, confidence, and kill criterion. report.html shows the conclusion first, then the idea and every round, with a Print / Save PDF button.

final.md 記下結論、信心分數與停止條件；report.html 先放結論，接著是點子與每一輪，並附上 Print / Save PDF 按鈕。

Make files, then write

先開檔，再動筆

Two scripts make the folder and the report. Claude writes the research and the arguments.

兩個腳本負責資料夾和報告。研究和論點是 Claude 寫的。

  1. Ask Claude

    Ask Claude Code to test an idea, or call /idea-review-panel. The default is 5 rounds.

  2. Make folder

    A script makes a dated folder with an input file, one file per round, and a final file.

  3. Fill rounds

    Research goes into the input file or the first round. Claude then fills the blank headings.

  4. Build report

    A second script joins them into the report.

  • default length is 5
  • headings start blank
  • the report is assembled
The folder name keeps the first eight words of the idea, cut at 64 characters. If none of those words can be used, the name is idea-review.

  1. 請它審

    請 Claude Code 驗證一個點子，或呼叫 /idea-review-panel。預設 5 輪。

  2. 建資料夾

    腳本會開一個帶日期的資料夾，放輸入檔、各輪一個檔和結案檔。

  3. 把各輪寫上

    研究寫進輸入檔或第一輪。Claude 再把空白標題填上。

  4. 接成報告

    另一支腳本再把這些檔接成報告。

  • 預設長度是 5 輪
  • 標題一開始是空的
  • 報告是接起來的
資料夾名稱取點子的前 8 個詞，最長 64 字。一個詞都取不出來時，名字是 idea-review。

No key, keep going

沒金鑰也繼續

The smoke test only reports which CLIs answer. The skill asks once, then continues.

短句探測只回報哪些 CLI 有回答。金鑰問一次，然後就繼續。

  1. Check CLIs

    Each CLI gets one short line and 45 seconds to answer.

  2. Ask once

    If the CLI exists but doesn't answer, Claude says the account is not reachable, asks once, and goes on.

  • smoke sends one line
  • the key is asked once
  • the fallback is labeled
When there is no key, Claude plays that provider and labels it, for example Claude simulating Grok.

  1. 先問 CLI

    每個 CLI 只拿到一句短的，而且要在 45 秒內回答。

  2. 問完就走

    CLI 在卻沒回答時，Claude 會說測不到帳號，問一次就走。

  • 探測只送出一句
  • 金鑰只問一次
  • 退路會標明是誰
沒有金鑰時，Claude 會扮演那個提供者並標明，例如 Claude 模擬 Grok。

The folder stays put

檔案留在本機

The review folder never leaves the working directory. Smoke is the step that hands a CLI one line.

審查資料夾留在工作目錄。探測的時候，只交一句話給 CLI。

  1. Local files

    check_providers.py prints which key names are set, and never prints the values.

  2. Send one line

    With --smoke the prompt is Reply OK only, and the wait stops at 45 seconds.

  • values are not printed
  • no smoke means no prompt
  • Python standard library only
  • on disk
  • one line
  • the CLI
Without --smoke the script reports the CLIs and key names it found, and it sends nothing. The scripts import only the Python standard library.

  1. 留在磁碟

    check_providers.py 只印有設定的金鑰名稱，不印那些值。

  2. 送一句出去

    加上 --smoke 時，送出的是 Reply OK only，最多等 45 秒。

  • 金鑰的值不會印
  • 沒加探測就不送
  • 只用標準程式庫
  • 磁碟上
  • 那一句
  • 某個 CLI
沒加 --smoke 時，只回報找到的 CLI 和金鑰名稱，不送出提示。這些腳本只用 Python 標準程式庫。

Verdict comes first

結論放最前面

The script only joins files that are already on disk. It never calls a model.

腳本只把磁碟上已有的檔接起來。它不會去呼叫模型。

  1. Read the files

    It needs the input and the verdict before it joins anything.

  2. Put verdict first

    The verdict comes first, then the input, then each round.

  3. Write Markdown

    The joined text is saved as a Markdown report.

  4. Write the page

    The same words are also written out as a printable web page.

  • the parts stay in that order
  • two files are written
  • the page is Traditional Chinese
Tables and code blocks stay plain text on that page.

  1. 先把檔讀進

    輸入和結論都要在，才接得起來。

  2. 結論排最前

    結論排最前面，接著是輸入，然後每一輪。

  3. 存成報告檔

    接好的文字存成一份 Markdown 報告。

  4. 寫成網頁

    同一份內容也寫成可以列印的網頁。

  • 這三段順序固定
  • 會寫出兩個檔
  • 網頁是繁體中文
表格和程式碼區塊在那個頁面上還是純文字。
What you get你會得到什麼

A decision aid, not encouragement.

要的是決策依據，不是加油打氣。

Most of the skill is instructions in SKILL.md. Three small Python scripts handle the mechanical parts: the provider check, the review folder, and the report.

這個 skill 大部分是 SKILL.md 裡的指示；三個小型 Python scripts 負責機械性的部分：provider 檢查、review 資料夾與報告。

01

Four personalities, not four models

四種個性，不是四個模型

Optimist, Skeptic, Pragmatist, and Synthesizer are thinking styles, not departments, roles, or separate models. Each gets its own section in every round file.

Optimist、Skeptic、Pragmatist、Synthesizer 是思考風格，不是部門、職務，也不是不同的模型。每一輪的檔案裡，各自都有一段。

SKILL.md
02

Native when possible, labeled when simulated

能用原生就用，模擬時會標明

Codex, Gemini, and Grok are used through their CLIs when they answer. Otherwise Claude asks once for an API key; without one, it simulates that style and labels it, as in Claude 模擬 Codex (Claude simulating Codex).

Codex、Gemini、Grok 的 CLI 能回應時就直接使用；不行的話，Claude 只問一次 API key，沒有 key 就模擬該風格並標示出來，例如「Claude 模擬 Codex」。

scripts/check_providers.py
03

Files you can keep

留得下來的檔案

Unless you ask for inline-only output, each review writes input.md, one file per round, final.md, report.md, and a static report.html you can print or save as PDF.

除非你只要在對話裡看結果，每次 review 都會寫出 input.md、每輪一個檔案、final.md、report.md，以及可以列印或存成 PDF 的靜態 report.html。

.idea-review/

Four styles, one session

四種風格同一輪

These are ways of thinking, not four model calls. The script writes the headings and stops.

這是四種想法，不是四次模型呼叫。腳本只寫標題就停。

  1. Optimist

    Write the strongest version of the idea, and why it could work.

  2. Skeptic

    Name a weak assumption, a hidden cost, or how it fails.

  3. Pragmatist

    Turn the argument into a test, and name the MVP and the constraints.

  4. Synthesizer

    Say where they disagree, and what evidence would change the call.

  5. Round file

    The script puts these four headings into every round file.

  • order is fixed
  • one Claude session
  • headings start empty
The order is Optimist, Skeptic, Pragmatist, then Synthesizer.

  1. 樂觀者

    寫出這個點子最強的版本，還有它為什麼可能做成。

  2. 懷疑者

    點出一個太樂觀的假設、一筆沒算到的成本，或它會怎麼失敗。

  3. 務實者

    把爭論收成一個小實驗，並寫下 MVP 和實際限制。

  4. 統合者

    寫出兩邊差在哪，以及什麼證據會讓結論改口。

  5. 寫進輪次檔

    腳本在每一輪的檔案裡放上這四個標題。

  • 四個標題順序固定
  • 同一次 Claude 填完
  • 標題一開始是空的
順序是樂觀者、懷疑者、務實者，最後是統合者。

Later rounds reply

後面的輪要回應

The rules are written down for Claude. None of the scripts check that they were followed.

規則寫給 Claude 看。三個腳本都不會檢查有沒有照做。

  1. First positions

    Round 1 asks for the strongest case, the likely failure, a small test, and the tension.

  2. Cite someone

    After round 1, each personality must use a point another personality already made.

  3. Pick a side

    In the last round each personality commits to build, validate, pivot, or stop.

  • stops at 5, 12, or 16
  • scripts do not enforce it
  • the verdict uses other words
The final file records one verdict: build it, test it first, change direction, or drop it. It also records a confidence from 0 to 100 and a condition for stopping.

  1. 先各自表態

    第 1 輪要最強的說法、可能的失敗、最小的測試，和決策上的張力。

  2. 要引用別人

    第 1 輪之後，每個個性都要用到另一個個性已經寫過的一點。

  3. 最後要選邊

    最後一輪，每個個性要在 build、validate、pivot、stop 裡選一個。

  • 停在 5、12 或 16
  • 腳本不會強制檢查
  • 結論欄用的是另一組詞
結案檔記下一個結論：直接做、先測試、改方向，或放棄。另外有 0 到 100 的信心，和一條停止條件。
Design decisions設計取捨

Why it is shaped this way.

為什麼這樣設計。

01

Personalities instead of model seats

用思考個性，而不是模型席位

The web version, ai-brainstorming, gives five models a seat each through OpenRouter. Inside Claude Code, four personalities need no extra accounts, and other providers are used only when their CLI answers or you give a key.

網頁版的 ai-brainstorming 透過 OpenRouter 讓五個模型各坐一席。在 Claude Code 裡，四種思考個性不需要額外帳號，其他 provider 只在 CLI 能回應或你提供 key 時才會用到。

02

Ask once, then keep going

只問一次，然後繼續

A missing key should not stall a review. The skill asks one time whether you have a key and whether you prefer the provider's own API or OpenRouter; if none is given, it continues on Claude simulation instead of waiting.

缺 key 不該讓整個 review 停下來。這個 skill 只會問一次：你有沒有 key、要用原廠 API 還是 OpenRouter；沒有提供，就改用 Claude 模擬繼續跑，不會一直等。

03

Conclusion first, rounds after

先放結論，再放過程

The report opens with the verdict and kill criterion, so a reader who stops after one page still gets the decision. The rounds follow as the supporting record.

報告一開頭就是結論與停止條件，只看第一頁的人也拿得到決定；每一輪的內容放在後面，當作佐證紀錄。

Start here從這裡開始

Clone it, then ask.

裝好之後，直接問。

You need Claude Code and Python 3.8+. A running session picks up the skill without a restart; if ~/.claude/skills/ is new, run /reload-skills once.

需要 Claude Code 與 Python 3.8 以上。正在執行的 session 不必重開就會載入這個 skill；如果 ~/.claude/skills/ 是新建立的，執行一次 /reload-skills。

Full install guide英文 README 完整說明中文安裝說明中文安裝說明

Install

安裝

Clone into your personal skills folder. For one project only, use that project's .claude/skills/ instead. On Windows, the README has a PowerShell version.

clone 到個人的 skills 資料夾；只想用在單一專案的話，改放到該專案的 .claude/skills/。Windows 的 PowerShell 版本請看 README。

git clone \
  https://github.com/teddashh/claude-idea-review-skill.git \
  ~/.claude/skills/idea-review-panel

Ask

開口問

The skill triggers from its description, or call it with /idea-review-panel. Ask for 12 or 16 rounds for a deeper review.

skill 會依描述自動觸發，也可以用 /idea-review-panel 直接呼叫。想更深入，就指定 12 或 16 輪。

幫我驗證這個點子：<你的 idea>
Validate this idea: <your idea>

Read the verdict

看結論

Claude replies with the verdict and the file paths. Open report.html in the .idea-review/ folder to print it or save it as PDF.

Claude 會回覆結論與檔案路徑。打開 .idea-review/ 裡的 report.html，就能列印或存成 PDF。

Honest status誠實的現況

Where it stands.

目前的樣子。

Built over six commits from June 14 to 16, 2026, as an open-source skill for the 自由工坊 Discord community and the 個人 IP 陪跑計畫. The skill has not changed since, and there are no releases; install from main.

這是 2026 年 6 月 14 日到 16 日、共 6 個 commit 寫成的開源 skill，為「自由工坊」Discord 社群與個人 IP 陪跑計畫而做。之後 skill 本身沒有再更新，也沒有 release，請直接從 main 安裝。

Checked on 2026-09-30

2026-09-30 實測

  • init_review.py creates the timestamped folder, input.md, one file per round, and final.md; --rounds accepts only 5, 12, or 16
  • init_review.py 會建立帶時間戳記的資料夾、input.md、每輪一個檔案與 final.md；--rounds 只接受 5、12、16
  • render_report.py writes report.md and a static report.html with a Print / Save PDF button and a page break between parts
  • render_report.py 會寫出 report.md 與靜態的 report.html，附 Print / Save PDF 按鈕，各部分之間會分頁
  • check_providers.py finds codex, agy or gemini, and grok on PATH, and lists key variable names without their values
  • check_providers.py 能在 PATH 上找到 codex、agy 或 gemini，以及 grok，而且只列出 key 變數名稱、不印出值
  • --smoke counts a CLI as answering only when it exits 0 with output; checked with stub CLIs
  • --smoke 只有在 CLI 以結束碼 0 結束、而且有輸出時才算通過；已用 stub CLI 測過
  • Chinese ideas keep their characters in the folder name
  • 中文點子的資料夾名稱會保留中文字
  • The scripts use only the Python standard library
  • scripts 只用 Python 標準函式庫

Limits and not yet

限制與尚未完成

  • The review is instructions in SKILL.md, not code, so its structure and quality depend on Claude following them
  • review 是 SKILL.md 裡的指示，不是程式碼，結構與品質取決於 Claude 有沒有照做
  • No API client: if you provide a key, how the API gets called is left to Claude in that session
  • 沒有 API 用戶端：提供 key 之後怎麼呼叫 API，交給當下 session 裡的 Claude
  • A failed agy smoke test is inconclusive, because agy -p can exit 0 with empty output when piped
  • agy 測試失敗不能下定論，因為 agy -p 在管線中可能以結束碼 0 結束卻沒有輸出
  • The report renderer does not support tables or code blocks, and always marks the HTML as zh-Hant
  • 報告產生器不支援表格與程式碼區塊，HTML 也一律標示為 zh-Hant
  • No tests, CI, or releases
  • 沒有測試、CI 或 release
Skill nameSkill 名稱
idea-review-panel
License授權
MIT
Requires需求
Claude Code · Python 3.8+
Optional選配
Codex · Antigravity / Gemini · Grok
Last code change最後程式更新
2026-06-16
Verified查核日期
2026-10-06

More from Ted Huang. Every public project has a page like this one, in English and Traditional Chinese.

Ted Huang 的其他作品。每個公開專案都有一頁像這樣的中英雙語介紹。

All projects →全部專案 →