chatgpt-consult-loop
NewUse when Worker Agent needs persistent ChatGPT consultant for plan review, stuck diagnosis, or completion validation. Multi-turn consult loop on same ChatGPT web conversation — not single-shot Q&A.
Summary
This skill enables a Worker Agent to maintain a persistent consult loop with ChatGPT Web for multi-turn plan review, stuck diagnosis, and completion validation.
- It leverages the same ChatGPT conversation context to provide coherent guidance across task stages, reducing errors and improving task success rates.
Overview
ChatGPT Persistent Consult Loop
Worker Agent 與 ChatGPT Web 的持久化顧問迴圈。不是一次性問答,而是同一任務中持續互動的顧問 / reviewer / 輕量級領導。
Overview
Worker Agent 在執行任務時,可以在關鍵節點(計畫審查、卡關診斷、完成驗收)回到同一個 ChatGPT 對話視窗請教顧問。ChatGPT 保留前文脈絡,能提供有連貫性的指導。
角色分工:
- •使用者:任務委託者 / Final Authority
- •Worker Agent:實際操作者,可自主判斷與執行。指的是 Hermes Agent、OpenCode、OpenClaw、Claude Code、Codex CLI 等 agent 軟體;底層模型可使用 DeepSeek V4 Pro、Qwen、MiMo 等便宜模型。
- •ChatGPT Web:Persistent Consultant / Reviewer,審查計畫、診斷卡關、驗收完成
When to Use
- •任務需要「第二意見」或方向校正
- •Worker 要執行多步驟任務,需要在關鍵節點被 review
- •任務有風險,需要在執行前被挑錯
- •Worker 連續失敗需要診斷
- •任務完成前需要外部驗收
When NOT to Use
- •簡單單步驟任務(直接做就好)
- •任務不需要外部意見
- •ChatGPT 頁面不可用或未登入
Prerequisites
- Chrome 已開啟並登入 chatgpt.com
- chrome://inspect/#remote-debugging 已打勾
- Chrome DevTools MCP 已連接(
npx -y chrome-devtools-mcp@latest --autoConnect --no-usage-statistics) - Agent 已載入此 skill
Standard Flow
Step 0: Bootstrap(每個新任務的第一輪)
每個新任務開始時,先在 ChatGPT 對話中送出 bootstrap 訊息。
讀取 templates/bootstrap.md 的內容,透過 MCP chrome-devtools 送出到 ChatGPT 輸入框。
等待回覆完成,確認 ChatGPT 理解角色。
Step 1: Plan Review(任務開始前)
Worker 產生初始理解與計畫後,讀取 templates/plan_review.md,填入:
- •原始任務
- •Worker 的理解
- •初始計畫
- •已知限制
- •風險或不確定性
送出到同一個 ChatGPT 對話,等待 review 回覆。
Step 2: Stuck After Two Failures(連續失敗時)
如果同一 subgoal 連續失敗兩次,立即停下來。
讀取 templates/stuck_after_two_failures.md,填入:
- •原始任務 / 目前 subgoal
- •兩次嘗試的過程與錯誤
- •目前狀態
- •具體問題
Step 3: Completion Review(完成前)
Worker 認為完成後,讀取 templates/completion_review.md,填入:
- •原始任務 / 完成摘要
- •實際改了什麼
- •測試結果
- •剩餘不確定性
Step 4: Adoption Rule(採用 / 不採用建議)
收到 ChatGPT 建議後,Worker 不可盲目照做。應產生:
ChatGPT 建議:...
Worker 採用:...
Worker 不採用:...
理由:...
下一步:...MCP Interaction Pattern
所有與 ChatGPT 的互動透過 Chrome DevTools MCP:
mcp_chrome_devtools_click(uid)— 點擊輸入框mcp_chrome_devtools_type_text(text=prompt, submitKey="Enter")— 輸入 + Entermcp_chrome_devtools_evaluate_script(function=waitForReply())— 等待回覆穩定- 儲存結果到
logs/
waitForReply 穩定判斷
async () => {
const waitForReply = () => new Promise((resolve) => {
let lastContent = "";
let stableCount = 0;
let attempts = 0;
const check = () => {
attempts++;
const msgs = document.querySelectorAll(
'[data-message-author-role="assistant"]'
);
if (msgs.length === 0) {
if (attempts > 60) resolve("TIMEOUT_NO_REPLY");
else setTimeout(check, 1500);
return;
}
const last = msgs[msgs.length - 1];
const content = last.innerText || "";
if (content && content === lastContent && content.length > 20) {
stableCount++;
if (stableCount >= 3) { resolve(content); return; }
} else {
stableCount = 0;
lastContent = content;
}
if (attempts > 60) resolve(lastContent || "TIMEOUT");
else setTimeout(check, 1500);
};
setTimeout(check, 3000);
});
return await waitForReply();
}Logging
每輪 consultation 都要存檔到 logs/:
logs/
├── 001_bootstrap.md
├── 002_plan_review.md
├── 003_stuck_diagnosis.md
└── 004_completion_review.mdConversation Persistence Rules
- 同一任務用同一個 conversation:不要每次開新對話
- 檢查 conversation URL:送出前確認 URL 包含同一個 conversation ID
- 斷線重連:記住 conversation URL,斷線後導航回去
- Context overflow:超長對話定期請 ChatGPT 摘要共識
Common Pitfalls
- 意外開新對話 — 最大風險。每次送訊息前檢查 URL 是否一致。
- 回覆穩定誤判 — ChatGPT 偶爾分段送出,穩定次數閾值至少 3。
- 輸入框狀態異常 — 送出前驗證 editor 存在且可聚焦。
- 盲目照做建議 — 每次重要建議後必須產生 adoption judgment。
- 無限延伸任務 — ChatGPT 的額外優化建議非阻塞項目,先列入後續。
Verification Checklist
- •[ ] Chrome 已開啟 chatgpt.com 並登入
- •[ ] chrome://inspect/#remote-debugging 已打勾
- •[ ] MCP chrome-devtools 已連接
- •[ ] Bootstrap 已送出並確認回覆
- •[ ] 每輪 consultation 都存到 logs/
- •[ ] Worker 不盲目照做 ChatGPT 建議
- •[ ] 完成前經過 completion review
Install & Usage
mkdir -p .claude/agentsAdd the configuration to .claude/agents/chatgpt-consult-loop.md
@chatgpt-consult-loopUse Cases
Usage Examples
/chatgpt-consult-loop I need to set up a CI/CD pipeline for a monorepo. Review my plan before I start.
/chatgpt-consult-loop I've failed twice trying to configure the database migration. Help me diagnose what's wrong.
/chatgpt-consult-loop I think the refactoring is complete. Can you validate the changes meet all acceptance criteria?
Security Audits
Frequently Asked Questions
What is chatgpt-consult-loop?
This skill enables a Worker Agent to maintain a persistent consult loop with ChatGPT Web for multi-turn plan review, stuck diagnosis, and completion validation. It leverages the same ChatGPT conversation context to provide coherent guidance across task stages, reducing errors and improving task success rates.
How to install chatgpt-consult-loop?
To install chatgpt-consult-loop: create the agents directory (mkdir -p .claude/agents), then add the config to .claude/agents/chatgpt-consult-loop.md. Finally, @chatgpt-consult-loop in Claude Code.
What is chatgpt-consult-loop best for?
chatgpt-consult-loop is a agent categorized under General. It is designed for: code-review, agent. Created by castlen3.
What can I use chatgpt-consult-loop for?
chatgpt-consult-loop is useful for: Review a multi-step implementation plan before execution to catch potential issues early.; Diagnose the root cause after two consecutive failures on the same subgoal during a complex task.; Validate that a completed task meets all requirements and edge cases before final delivery.; Get a second opinion on architectural decisions or code design choices during development.; Request guidance on how to proceed when the worker agent is uncertain about the next step.; Perform a final quality check on generated code or configuration before merging..