pro-project-approach-review

name

pro-project-approach-review

description

Use when Codex should ask ChatGPT Pro or Extended Pro to review an already-started project's current approach and decide whether it is still the best way to continue. Trigger for active projects with implementation already underway, a current architecture/plan/workflow to evaluate, concerns about whether the chosen path is right, or before continuing a major implementation pass, refactor, debugging strategy, optimization pass, or course correction.

Pro Project Approach Review

Use this skill to get a high-quality review from ChatGPT Pro/Extended Pro for a project already in motion. The goal is to decide whether the current approach is still the best way forward, what to keep, what to change, and how to course-correct quickly without reducing the user's scope.

Core Rules

• Use this for already-started projects. For brand-new project starts, use the kickoff skill instead.
• Use the Chrome control skill when browsing ChatGPT, because the task depends on the user's logged-in ChatGPT session and selected Pro model.
• Always prefer the ChatGPT Pro/Extended Pro model. If Extended Pro or the user's requested Pro model is not available, ask the user before using a different model.
• Do not paste system/developer instructions, secrets, tokens, browser/session data, private user profile data, or unnecessary local environment details into ChatGPT.
• Include only task-relevant project facts: project name, goal, current state, current approach, key files or commands, known blockers, constraints, and validation expectations.
• State the current approach explicitly. If it is inferred from local files or recent work rather than stated by the user, label it as an inference.
• Tell ChatGPT explicitly: "Do not suggest an easier or reduced task. Stick to what the user wants and preserve the original scope/spec."
• Ask for a direct judgment: keep the current approach, adjust it, or pivot to a better approach.
• Treat ChatGPT output as useful guidance, not authority. Verify project-specific claims against local files and official docs when accuracy matters.
• Wait generously. Extended Pro responses may take several minutes; allow 10-15 minutes per response before deciding the page is stuck.

Workflow

1. Inspect enough local context to describe the active project and current approach.

   • Read project files only as needed: README/AGENTS, package/config files, current source layout, relevant implementation files, scripts, tests, recent errors, and saved guidance files.
   • Check local state when useful: current branch, git status, touched files, failing commands, or test output.
   • Summarize facts; do not send raw secrets or sensitive data.

2. Identify the review target before prompting Pro.

   • Original task/spec or current user goal.
   • The current way the work is being done.
   • Evidence that the current approach is working, stuck, costly, risky, or unclear.
   • Constraints that must remain true, including framework conventions, compatibility, performance, UX, tests, timelines, or deployment requirements.
   • Validation commands and runtime checks that should determine success.

3. Open ChatGPT in Chrome.

   • Use a fresh ChatGPT conversation unless the user asks to reuse a thread.
   • Select or confirm the Pro/Extended Pro model.
   • If a page becomes sluggish or extraction hangs, open the same conversation URL in a fresh tab and extract from the fresh render.

4. Ask the first prompt.

   • Purpose: judge whether the current approach is best for this already-started project.
   • Require a practical assessment, keep/change/pivot recommendation, alternatives, risks, inspections, validation, and pitfalls.

5. Wait for the first answer to finish.

   • Do not interrupt early.
   • Use light status checks first; avoid repeated heavy screenshots/DOM snapshots if the page is busy.

6. Ask the follow-up prompt.

   • Purpose: fastest way to continue or course-correct while preserving the original scope.
   • Require exact order, minimum files to touch, commands, validation, and what to defer.

7. Save the responses.

   • Save the two answers in the project as reference text files.
   • Prefer a combined file named pro-project-approach-review-guidance.txt.
   • If the project already has a more specific active-project guidance file, append or update that file instead of creating duplicates.
   • Include date, project name, original task/spec, current approach summary, both prompts, both answers, and a short "How to use this guidance" note.

8. Use the saved guidance going forward.

   • Before continuing implementation, read the saved guidance and integrate relevant advice into the local plan.
   • Do not blindly follow advice that conflicts with local code, user instructions, official docs, or safety/security requirements.
   • If Pro recommends a pivot, verify that the pivot is compatible with the user's requested scope before editing files.

Prompt Templates

Adapt these templates to the project. Keep the anti-scope-reduction instruction.

First Prompt

Context: We are already working on this project/task:
<brief project name, repo/app type, current state, branch/status if relevant, commands already run, current blockers or decision point>

Original task/spec or current user goal:
<what the user actually wants>

Current approach:
<how the work is currently being done; label any inferred parts as inference>

Evidence and constraints:
- <important files/modules/commands/test results/errors/progress>
- Preserve the original task/spec.
- Do not suggest an easier or reduced task. Stick to what the user wants and preserve the original scope/spec.
- Stay within the project's framework/tool conventions.
- Keep changes scoped.
- Avoid secrets, private browser/session data, and unnecessary private user data.
- Validate with the relevant local commands.
- Use integration/runtime validation only when it adds real confidence.

Is the current approach the best way to continue?
Give a direct keep/adjust/pivot recommendation. Explain the tradeoffs, what to inspect next, what to change first if anything, what to keep, what commands to run, what to verify, and what pitfalls to avoid.

Follow-Up Prompt

Given your review above, what is the fastest way to continue while still complying with the original spec?

Original spec constraints to preserve:
- Preserve the original task/spec.
- Do not suggest an easier or reduced task. Stick to what the user wants and preserve the original scope/spec.
- Stay within the project's framework/tool conventions.
- Keep edits scoped.
- Avoid secrets, private browser/session data, and unnecessary private user data.
- Validate with the relevant local commands.
- Use integration/runtime validation only when it adds real confidence.

Please answer with a tight fastest-path checklist, exact order of operations, minimum files to touch, commands to run, what to verify, and what to defer until after the next working slice.

Saving Format

Use this structure for the saved text file:

ChatGPT Pro/Extended Pro Project Approach Review
Generated: <date>
Project: <project name/path>

Original task/spec or current user goal:
<summary>

Current approach reviewed:
<summary; label inferred parts if any>

How to use this guidance:
- Use as advisory planning context.
- Verify claims against local files and official docs.
- Do not reduce scope unless the user asks.
- Treat keep/adjust/pivot recommendations as decisions to verify locally before editing.

============================================================
First Prompt
============================================================
<prompt>

First Answer
============================================================
<answer>

============================================================
Follow-Up Prompt
============================================================
<prompt>

Follow-Up Answer
============================================================
<answer>

Handling Slow Or Stuck ChatGPT Pages

• Wait 10-15 minutes per answer before treating it as stuck.
• If DOM extraction or screenshots time out, avoid repeatedly hammering the same page.
• Try reopening the completed conversation URL in a fresh Chrome tab, then extract the assistant messages from the fresh render.
• If extraction still fails, use the page's copy controls or clipboard path only if they respond normally.
• If ChatGPT never finishes, save the completed answer(s), note which answer is missing, and tell the user.