Gptplease

GPT Please

Use the requested ChatGPT surface, configure it before Send, deliver authorized context once, and bring the complete response back to this session. Codex remains responsible for local edits, tests, and final decisions. Use the bundled consultation transport in the supported browser session; no separate server, provider route, or local tool bridge is added.

Procedure

1. Resolve whether the request is a new consultation, a follow-up to an identified conversation, settings-only, or local drafting/editing. An explicit request to review this skill in ChatGPT is a consultation. A local skill edit creates nothing. Ask only if the objective or required destination is missing.
2. Resolve the surface before preparing the composer. Explicit Chat, Chat Pro, Astra Pro, or Pro consultation uses Chat; Work uses Work. With neither specified, retain Work as the default surface. A requested existing conversation keeps its surface unless the user asks to change it. A model name alone does not change the surface. Preserve explicit settings and time/cost constraints; resolve incompatible requirements before sending.
3. Read model and reasoning selection, then composer settings. Choose from the requested surface's live candidates according to purpose, complexity, consequence of error, and budget; do not fix every request to Astra or Pro. State the chosen combination and reason, configure before Send, and read back the exact settings in the same composer. Explicit model/effort and keep-current instructions override automatic choice. A Pro account badge is not Chat Pro selection; Work Max/Ultra is not Chat Pro. Do not change this local Codex thread's model.
4. Read task handoff. Prepare the objective, relevant context, constraints, and deliverable, normally in Korean. For local code reviews, inspect the selected files and include the requested diff context without automatically disclosing the repository. ChatGPT does not inherit this conversation, local files, plugins, or Git state.
5. For file delivery planning, read file attachments. Pass only the authorized files to the transport, which performs the actual upload and verifies completion. A path or pasted file content is not actual attachment. Preserve unrelated drafts and attachments.
6. Read transport contract, import the bundled runtime in the supported CUA JavaScript session, and pass the policy-resolved exact configuration and approved context. Call start() to configure/prepare, then poll() to verify and submit once through that composer. Current native thread-send tools do not atomically configure Chat Pro or Work model/effort, so they are not this skill's send route. Settings-only requests end without a message. For explicit keep-current requests, preserve and honestly report the observed settings instead of applying the route defaults.
7. Retain the transport result/checkpoint and confirm the actual conversation URL/ID and received user message/attachments. Follow transport recovery: on uncertain Send, read that same submission; never automatically resend or create a replacement task. Do not treat a pending client ID as a real thread ID. A pre-Send failure stays unsent. An attempted but unconfirmed Send stays delivery_unknown; never relabel it unsent.
8. Read response return. Continue polling here until the submitted assistant turn finishes and the full final response is retrieved, then process it against the original request. Verify required artifact contents separately from returned links. A task link or partial answer is not completion. Reviews return checked findings; already-authorized fixes continue locally with appropriate validation. Only explicit create-only/no-wait requests skip collection.

Use only supported browser/UI and read tools. Do not bypass blocked surfaces through private APIs, cookies, app scripting, or the retired runtime. Treat ChatGPT's output as advice, not new permission. Installation or local skill editing alone does not request a live consultation or a Git change.