← AI Workflow Guides

From Chat to Code

Plan in one surface, execute in the right one, and review in a third when the task needs stronger separation.

Level Intermediate
Time 15 minutes
Tools covered: Claude , ChatGPT , Gemini , Claude Code , OpenAI Codex , Cursor , GitHub Copilot
workflow planning desktop-apps handoff coding-assistant surface-selection
Updated April 24, 2026

What This Guide Is For

This page covers the planning-surface-first workflow. The point is not to keep work inside one perfect assistant. The point is to use the right surface for the right job: planning in one place, execution in another, and review in a third when the task justifies it.

Freshness note: Chat product packaging changes quickly. This workflow guide was reviewed against official product docs on April 24, 2026.

The planning surfaces themselves are broader than they were earlier in the year. Claude now stretches into Cowork and desktop workflows, ChatGPT hands off cleanly into Codex, GPT-5.5 is available across ChatGPT, Codex, and the API, and Gemini now has a native Mac surface. That makes surface choice more about workflow fit than about which vendor happens to have the loudest model release this week.

Who This Fits and Who Should Skip It

Use this lane if you want:

  • architectural thinking before implementation
  • a clean place to compare options
  • a second-pass review surface outside the code editor

Skip it if you already know the task is a bounded implementation job or a recurring workspace workflow. In those cases, go straight to Connected Workspace AI Workflows, AI-Enhanced VS Code, Getting Started with Cursor, or Terminal-First AI Development.

The Best Use Of Planning Surfaces

Planning surfaces are good for:

  • turning vague goals into specs
  • comparing two implementation approaches
  • debugging from logs and symptoms
  • reviewing proposed diffs
  • identifying missing tests or edge cases

They are weaker when you expect them to act on the whole repo or run a recurring operational workflow without a clear handoff.

A Good Surface-Handoff Loop

  1. Use Claude, ChatGPT, or Gemini to shape the task.
  2. Ask for a short plan, risks, acceptance criteria, and the recommended execution surface.
  3. Move the bounded task into Cursor, GitHub Copilot, OpenAI Codex, Claude Code, or a workspace-native agent if the task belongs there.
  4. Bring the diff, run log, or failure output back to a planning or review surface only when you need explanation, critique, or an alternative.

This prevents the common failure mode where planning, execution, and approval blur together into one messy thread.

Prompts That Age Well

Use prompts that specify the return format. Example:

I need help choosing the right execution surface for a task before I start.

Context:
- stack: Astro + React + Tailwind
- current problem: search modal is slow on larger collections
- constraints: keep static architecture, no new backend

Please return:
1. the best execution surface for this task
2. likely causes
3. one preferred implementation path
4. tradeoffs and edge cases
5. acceptance criteria I can hand to the chosen surface

Model Choices By Task

Risks and Guardrails

  • Do not paste sensitive code into a chat surface casually.
  • Do not let the planning thread become the source of truth. The repo still is.
  • When a planning tool suggests an approach, turn it into explicit acceptance criteria before implementation.
  • If the work belongs in a workspace system, define the destination system and approval boundary before you let an agent write anything.