Start here
Choose your tool, define a developer task, and verify the result.
This guide covers Claude and Claude Code, ChatGPT and Codex, and Gemini and Gemini CLI. Start with a small real problem whose result you can check.
1. Choose where to do the work
Use a chat interface to reason from code, selected sources, or data and prepare a reviewable task. Use a coding agent when it should work with your repository and run agreed checks. Use the corresponding APIs for your own programs.
A model name does not tell you which files, tools, permissions, or usage allowance the interface has. Check the active account access too. Access and models
2. Define the behavior you want
The search field loses focus when results update. Find the cause and reproduce the failure. After the fix, typing, deleting, and keyboard navigation must work. Preserve the existing search rules. Show the diff and the checks actually run.
Supply the relevant code, reproduction, and known constraints. Ask the tool to identify missing requirements. Material unresolved design choices need a concrete plan before implementation.
3. Maintain context for the actual product
Project instructions, AGENTS.md, CLAUDE.md, and GEMINI.md have different loading and scope rules. Each tool guide explains its own setup. Keep goals, decisions, sources, and open questions concise and verifiable.
Understand context, retrieval, and compaction
4. Check the result against the original
Read the diff, run the relevant checks, and inspect the changed behavior. Open sources before relying on a technical claim. Confident wording or a second model's agreement does not replace that evidence.
5. Automate a repeatable handoff
When inputs, outputs, and checks are clear, turn the workflow into a pipeline recipe. Combine tools through reviewable artifacts and retain human decisions at meaningful boundaries. Compare the combination with your existing method.
Workflows and executable examples · Quick fixes for common problems