Chatgpt Webapp Delegator

ChatGPT WebApp Delegator

Overview

Use ChatGPT WebApp as a delegated reasoning surface through the Chrome plugin. Keep local context, source access, task memory, and final judgment in Codex; inject only the relevant context into ChatGPT WebApp, then evaluate and iterate before answering the user.

Core Workflow

1. Decide whether delegation is worthwhile. Use ChatGPT WebApp when it is likely to add material value: deep research, high-token synthesis, strategic critique, ideation breadth, long-form drafting, complex reasoning, or a second pass from another strong model. Stay local when Codex has better tool access, when the task is small, or when pasting the needed context would be unsafe.

2. Build a bounded context packet. Keep the complete working memory in Codex. Send ChatGPT WebApp only what it needs: the task, requested output shape, relevant excerpts or facts, constraints, definitions of success, known unknowns, and instructions about what not to assume. Read references/delegation-prompts.md when composing the first prompt or an iteration prompt.

3. Open ChatGPT WebApp with Chrome. Use the Chrome plugin/skill, claim a relevant existing ChatGPT tab if appropriate, or open https://chatgpt.com. If login, CAPTCHA, payment, quota, or account-selection blocks appear, pause for the user. Before typing or uploading sensitive, proprietary, private, regulated, or personal data, get explicit user approval because this transmits data into ChatGPT WebApp.

4. Select the strongest available model or mode at runtime. Inspect the live ChatGPT WebApp model/mode picker; do not hardcode model names, rankings, selectors, or labels. Interpret "best model" as the highest-capability option visible in the user's account for the task, unless the user prioritizes speed or quota conservation. If the apparent best option is unavailable, use the strongest visible fallback and say so when it matters.

5. Use specialized ChatGPT modes when they fit. Prefer Deep Research for source-backed research, market/company scans, current-topic reports, and citation-heavy synthesis. Prefer Agent for multi-step browser or app workflows only when broader web/app agency is useful and the user has approved the scope. Prefer the strongest reasoning/general chat mode for strategy, critique, coding architecture, math, and complex synthesis. Use file or data-analysis capabilities only after confirming that the file contents are appropriate to upload.

6. Run the delegation. Start a new ChatGPT conversation unless the user asks to continue an existing one. Paste the context packet, ask for the desired output, and require explicit assumptions, uncertainty, and missing evidence. For long-running Deep Research or Agent tasks, monitor progress and capture the final answer plus any important cited sources, action trace, or caveats.

7. Evaluate and iterate. Treat ChatGPT WebApp output as draft material, not authority. Compare it against the user's request, Codex's local context, available source evidence, and the stated evaluation bar. Iterate when the answer is generic, misses constraints, invents facts, overclaims, ignores local evidence, or needs sharper framing. For source-sensitive, high-stakes, or current claims, verify outside ChatGPT WebApp when practical.

8. Integrate the final answer in Codex. Give the user Codex's synthesized answer, not a raw dump, unless they ask for the raw ChatGPT output. Mention the ChatGPT WebApp model or mode used when it is visible and relevant. Clearly separate verified facts from ChatGPT-provided suggestions or unverified inferences.

Safety Rules

• Do not paste secrets, API keys, credentials, session tokens, passwords, or private browser/profile data into ChatGPT WebApp.
• Do not paste full private documents, emails, transcripts, financial data, health data, legal documents, or regulated information unless the user explicitly approves that transmission.
• Prefer summaries and targeted excerpts over wholesale uploads. Do not upload local files unless the user approves the specific file and purpose.
• Do not let ChatGPT WebApp send messages, submit forms, make purchases, modify accounts, change files, or take external actions unless the user authorizes that specific action.
• If Chrome or ChatGPT WebApp is unavailable, say so and continue locally when possible.
• Finalize Chrome tabs according to the Chrome skill after browser work is complete.

Quality Bar

A good delegation run produces an answer that is better than either model working alone: ChatGPT WebApp contributes breadth, depth, or compute; Codex contributes local context, source discipline, user memory, judgment, and final editing.

Before answering, check:

• Did the delegation answer the user's real question?
• Did Codex preserve all relevant local context and constraints?
• Are claims, sources, and assumptions appropriately qualified?
• Is another iteration needed to close a concrete gap?