chatgpt-worker

name	chatgpt-worker
description	Use this skill to send a prompt to ChatGPT and get a response. Triggers when the user wants to query ChatGPT, use ChatGPT as an external AI worker, or when the web-ai-router directs a task to ChatGPT. Best for general QA, writing, coding, and creative tasks. Requires login — session is persisted in Chrome profile.

ChatGPT Worker

Send a prompt to ChatGPT via browser and return the response.

Environment

• URL: https://chatgpt.com
• Browser: Chrome on localhost:9222 (managed by local-browser-tools)
• Login: Required — session persisted in ~/.cache/browser-tools profile
• Tools: browser-start.js, browser-nav.js, browser-eval.js, browser-screenshot.js

---

Input

prompt: string        # The text to send to ChatGPT
task_type: string     # qa | summary | translation | rewrite | extraction | comparison | brainstorming | analysis

---

Output

Always output this JSON block first:

{
  "site": "chatgpt",
  "task_type": "<task_type>",
  "status": "success | partial | error",
  "answer": "<extracted response text>",
  "retry_count": 0,
  "elapsed_ms": 0,
  "error_reason": null
}

Then follow with a short human-readable summary (2–3 sentences max).

---

Critical Architecture Note

ChatGPT uses ProseMirror as its rich text editor. The real input element is a div.ProseMirror, NOT the <textarea> on the page. The textarea (wcDTda_fallbackTextarea) is a hidden accessibility fallback that React does not monitor for input. Writing to the textarea will never trigger the send button — it will always launch voice mode instead.

---

Workflow

Step 1 — Start browser

node ~/.pi/agent/skills/local-browser-tools/browser-start.js

Expected: ✓ Chrome already running on :9222 or ✓ Chrome started on :9222

If fails: return error status, stop.

---

Step 2 — Navigate to ChatGPT

node ~/.pi/agent/skills/local-browser-tools/browser-nav.js https://chatgpt.com

Take a screenshot and verify:

• Input area is visible
• Not showing a login wall

If login page appears: return error_reason: "login_required", stop.

---

Step 3 — Locate input box

ChatGPT's real input is a ProseMirror div, NOT a textarea.

Check:

!!document.querySelector('div.ProseMirror[contenteditable="true"]')

Fallback selectors in order:

1. div.ProseMirror[contenteditable="true"]
2. div[role="textbox"][aria-label="Chat with ChatGPT"]
3. div[contenteditable="true"][role="textbox"]

DO NOT use textarea — it is the wrong element.

If no selector works after 3 attempts: trigger retry.

---

Step 4 — Insert prompt

Use execCommand("insertText") — this is the correct method for ProseMirror editors. Do NOT set innerHTML or textContent directly as ProseMirror will not register the change.

(function() {
  const input = document.querySelector('div.ProseMirror[contenteditable="true"]')
    || document.querySelector('div[role="textbox"][aria-label="Chat with ChatGPT"]')
    || document.querySelector('div[contenteditable="true"][role="textbox"]');
  if (!input) return "NOT_FOUND";
  input.focus();
  input.innerHTML = "";
  const sel = window.getSelection();
  const range = document.createRange();
  range.selectNodeContents(input);
  sel.removeAllRanges();
  sel.addRange(range);
  document.execCommand("insertText", false, PROMPT_TEXT);
  return input.textContent;
})()

Verify returned value matches the intended prompt. If mismatch: retry.

Take a screenshot to confirm text appears in the input area.

---

Step 5 — Send

After Step 4 succeeds, the composer-submit-button-color button will behave as a send button. Click it:

document.querySelector('button.composer-submit-button-color')?.click()

After clicking, take a screenshot to confirm:

• The message bubble appears in the chat
• The input area is cleared

If the input area is NOT cleared: Step 4 likely wrote to the wrong element. Re-check Step 3 selector and retry from Step 4.

---

Step 6 — Wait for response

Poll every 3 seconds, maximum 120 seconds total. Do NOT use fixed sleep.

Use response length stability as the primary completion signal:

1. After sending, start polling response text length
2. Take a screenshot on the first poll where length > 0 — save as "thinking" screenshot
3. Treat response as complete when length is non-zero and unchanged across 2 consecutive polls — save as "response complete" screenshot

Response extraction for polling:

(function() {
  const msgs = document.querySelectorAll('[data-message-author-role="assistant"]');
  if (msgs.length > 0) return msgs[msgs.length - 1].innerText?.trim()?.length || 0;
  return 0;
})()

If 120 seconds pass with no stable response: trigger retry.

---

Step 7 — Extract response

(function() {
  const msgs = document.querySelectorAll('[data-message-author-role="assistant"]');
  if (msgs.length > 0) {
    return msgs[msgs.length - 1].innerText?.trim();
  }
  const candidates = document.querySelectorAll('[class*="markdown"], [class*="prose"]');
  if (candidates.length > 0) {
    return candidates[candidates.length - 1].innerText?.trim();
  }
  return "EXTRACTION_FAILED";
})()

Validate the result:

• Must contain text
• Must not equal the original prompt
• Must not be UI chrome text (e.g. "Ask anything", "ChatGPT can make mistakes")

If validation fails: trigger retry.

---

Retry Policy

• Maximum retries: 2
• On each retry: re-navigate to https://chatgpt.com before retrying
• After 2nd failure: return status: "error" with error_reason explaining which step failed

---

Selector Change Resilience

If a selector stops working:

1. Use browser-eval.js to list all [contenteditable] elements: Array.from(document.querySelectorAll('[contenteditable]')).map(e => ({tag: e.tagName, role: e.getAttribute('role'), aria: e.getAttribute('aria-label'), cls: String(e.className).slice(0, 80)}))
2. Identify the ProseMirror editor by its className or role="textbox"
3. Never use the textarea element — it is always the wrong target

---

Example Output

{
  "site": "chatgpt",
  "task_type": "qa",
  "status": "success",
  "answer": "TEST-CHATGPT",
  "retry_count": 0,
  "elapsed_ms": 14000,
  "error_reason": null
}