gpt-pro-context-handoff

name

gpt-pro-context-handoff

description

Use when Codex needs to get a deep ChatGPT/GPT Pro review, audit, proposal, implementation plan, risk assessment, or strategy analysis for any project by serving curated context through a temporary HTTPS text endpoint and interacting with chatgpt.com via Browser Use. Triggers include requests to package context for GPT Pro, ask Pro to analyze a codebase/system/document set, serve project context from a reachable host, monitor a Pro response, or follow up with Pro using browser automation.

GPT Pro Context Handoff

Core Rule

Lead with plain readable text over HTTPS, not ZIP upload and not ZIP-primary links.

ZIP files, tarballs, and large archives are backup artifacts only. The reliable handoff is a .txt or .md context file that GPT Pro can browse/fetch/read directly, with optional evidence links for deeper material.

Non-Negotiables

• Treat the skill as reusable and project-neutral. Do not include private project names, hosts, paths, credentials, conventions, metrics, strategy, incidents, or domain assumptions unless they are explicitly present in the current user request or current workspace being packaged.
• Use Browser Use for chatgpt.com interaction when the user asks to open, prompt, wait, inspect, or follow up in the browser.
• Do not tell GPT Pro not to browse unless the user explicitly asks that.
• Do not interrupt GPT Pro while it is thinking, browsing, fetching, or finalizing unless the user explicitly asks.
• Do not treat HTTP 200 on an archive as proof GPT Pro read it. Confirm by the visible response and, when available, server logs for the plaintext context.
• Do not make ZIP upload through the ChatGPT UI the plan. Native file pickers and clipboard file-paste are unreliable.
• If GPT Pro says it cannot fetch, unzip, or read an artifact, pivot to smaller plaintext mirrors immediately.
• If the user says not to ask for confirmation, proceed and avoid confirmation prompts.

Workflow

1. Reconstruct The Request

Before touching the browser, write down the task in operational terms:

• desired GPT Pro output;
• current project goal, domain, and constraints;
• what evidence GPT Pro needs;
• what is in scope and out of scope;
• what GPT Pro must not assume;
• any user-specific framing, terminology, priorities, deadlines, risk posture, or operating constraints.

Preserve the user's framing for the current domain. Do not silently replace a project-specific goal with generic advice, and do not turn an execution request into a broad strategy memo unless that is what the user asked for.

2. Build A Plaintext Context File

Create one primary file named like:

gpt-pro-context-YYYYMMDDTHHMMSSZ.txt

Make it self-contained enough that GPT Pro can start without downloading archives. Include:

• REQUEST: exact desired analysis/proposal format.
• PROJECT ORIENTATION: project purpose, repo/path, runtime, architecture, important components, and glossary terms GPT Pro needs.
• CURRENT STATE: current status, metrics, incidents, open blockers, deployment/runtime state, or other domain-specific facts relevant to the request.
• RECENT CHANGES: what changed since the last review and what was verified.
• EVIDENCE INDEX: HTTPS links to text mirrors first; optional archive links second.
• KEY FILE EXCERPTS: concise excerpts or full contents for the most relevant source, prompts, configs, docs, and test output.
• RAW OUTPUTS: command output, logs, data summaries, status snapshots, and other request-relevant evidence.
• QUESTIONS FOR GPT PRO: direct, ranked questions.
• CONSTRAINTS: what not to suggest, compatibility requirements, dependencies that must remain enabled, risk posture, timing, and decision criteria.

Choose evidence for the current project rather than following a fixed domain checklist. Good inputs often include architecture docs, README files, key source files, prompts, schema or API definitions, config excerpts, test output, logs, diffs, performance traces, incident timelines, screenshots rendered to text descriptions, and prior decisions.

Do not include secrets. Redact keys, tokens, private keys, session cookies, production credentials, proprietary customer data, and private personal data. Prefer summaries over dumping secret-bearing config files.

3. Serve Context Over HTTPS

Use a user-approved or workspace-appropriate HTTPS host. If no host is already available, create a random-token directory on a reachable host or local tunnel that can serve static plaintext over HTTPS.

Requirements:

• public HTTPS URL;
• random unguessable path token;
• Content-Type: text/plain; charset=utf-8 for .txt;
• Accept-Ranges: bytes for large optional artifacts;
• curl -I from local machine returns 200;
• curl of the plaintext URL returns readable text;
• unknown paths return 404.

Do not assume any host, path convention, account, repository, deployment environment, or organization-specific workflow.

Archive links may be listed below the text URL, but mark them optional.

4. Prompt GPT Pro

Open or reuse chatgpt.com with Browser Use. Select GPT Pro/Extended Pro if available.

Prompt shape:

Please browse/fetch and read this plaintext project context first:
<TEXT_CONTEXT_URL>

Use it as the evidence of record for <specific requested review/audit/proposal>.

If any linked artifact is unavailable, say exactly which URL failed and continue from the plaintext context. Use optional archive links only if you need rawer context.

Focus on <user goal>. Be direct, cite evidence from the context, identify missing evidence, and produce prioritized actions with expected impact, risk, effort, confidence, and verification steps.

If the project has strong constraints, state them explicitly in the prompt. Example:

Do not recommend a rewrite unless the context shows a concrete blocker that cannot be solved incrementally. Rank proposals by impact, effort, risk, confidence, and how each would be verified in this project.

5. Wait And Monitor

After sending:

• Do not send follow-up while Pro is still thinking unless the user asks.
• Monitor the visible state: thinking, browsing, finalizing, stopped, or completed.
• Monitor server logs for plaintext URL hits and follow-up evidence hits.
• If Pro only acknowledges and does not fetch/read, follow up once with: “Do not just acknowledge; actually fetch/read the plaintext URL now. If blocked, say the blocker first.”
• If Pro cannot fetch a large archive, provide the relevant content as a smaller .txt follow-up file, not another ZIP attempt.

6. Capture And Act

When Pro finishes:

• extract the proposal and key findings;
• separate recommendations that conflict with user constraints;
• ask/answer with the user’s corrected framing;
• write any accepted plan to a repo doc if requested;
• implement and verify only after the user’s instruction or when the request already includes execution.

If continuing the same Pro thread, use follow-up plaintext files for corrected data, diffs, or compact patch bundles.

Context File Template

# GPT Pro Context Handoff

Created UTC: <timestamp>
Project: <name/path/repo/host>
Primary goal: <goal>
Domain: <product/domain/system type>

## REQUEST
<exact output wanted from GPT Pro>

## USER CONSTRAINTS
- <constraint>
- <constraint>

## PROJECT ORIENTATION
<purpose, architecture, runtime, key components, glossary, important paths>

## CURRENT STATE
<status, metrics, blockers, deployment/runtime state, recent behavior, open risks>

## VERIFIED FACTS
- <fact with command/source>

## RECENT CHANGES AND TESTS
<git diff summary, rollout status, test commands/results>

## EVIDENCE INDEX
Primary plaintext context:
- <this URL>

Additional plaintext evidence:
- <URL>

Optional archives only if needed:
- <URL>

## KEY SOURCE AND PROMPT EXCERPTS
<paths + excerpts or full relevant files>

## RAW OUTPUTS
<CLI/API/DB/log outputs>

## QUESTIONS FOR GPT PRO
1. <question>
2. <question>

## FAILURE MODES TO CHECK
- <risk>
- <risk>

## OUT OF SCOPE
- <thing GPT Pro should not spend time on>

Recovery Patterns

• ZIP did not work: create gpt-pro-context-followup-<stamp>.txt with the exact missing content.
• Pro asks for a direct URL: provide only the plaintext URL first; archives below it.
• Pro gives generic advice: follow up with user constraints and ask it to re-rank proposals under those constraints.
• Pro stalls but UI still shows thinking/finalizing: wait. Do not interrupt.
• Server logs show no context fetch: ask Pro once to fetch the text URL explicitly; verify URL with curl.
• Server logs show archive fetches but answer lacks citations: ask Pro to cite exact sections from the plaintext context or say what it could not read.

Completion Checklist

• Primary context is plaintext and served over HTTPS.
• URL verified locally with curl.
• Browser prompt includes the plaintext URL and explicit output structure.
• Optional archives are clearly marked optional.
• GPT Pro was not interrupted while thinking.
• Response was captured and reconciled with user constraints.
• Temporary public server/files are left running only as long as needed.