Module 20 lesson 05

Model Selection Without Overpromising

What You Will Learn

By the end, learners can explain model selection without overpromising, ask Codex for focused help, review the result, and decide the next safe step.

Why It Matters

Model Selection Without Overpromising matters because Codex is strongest when you can describe the desired outcome, the current project context, the constraints, and the evidence that proves the work is done. Beginners do not need to memorize every command. They need enough literacy to steer the agent, spot risky changes, and ask for a safer next step.

Plain-English Explanation

Think of this lesson as one practical layer in the Codex shipping loop: understand the work, define a small change, let Codex inspect before editing, review the diff, test the result, and only then decide whether to publish or continue. If a feature is plan-dependent, rolling out, or different across the app, CLI, IDE extension, and cloud/web task surfaces, say so in the prompt and ask Codex to verify the current surface before assuming it can act.

Model Selection Without Overpromising belongs to the larger advanced codex cli, config, models, and local workflows workflow. Treat it as a practical decision point: what should Codex inspect, what should it avoid, what evidence proves success, and what human review is required before shipping?

Practical example: if your goal is "Model Selection Without Overpromising", ask Codex to return a short map of the relevant files, a one-step beginner exercise, and a review checklist before making changes.

Step-by-Step Tutorial

1. Define the workflow boundary: inputs, outputs, tools, permissions, and data that must stay private.
2. Ask Codex to identify what is documented, what is assumed, and what must be verified in the current surface.
3. Start with a read-only or mock version before allowing real actions.
4. Add approval points for network calls, file writes, credentials, deployments, and destructive commands.
5. Run a focused review: security, correctness, maintainability, and test coverage.
6. Create an evaluation or checklist that scores the output against the goal.
7. Document when the workflow should not run.

Copy/Paste Codex Prompt

You are helping me learn Model Selection Without Overpromising. First explain the concept in plain English. Then inspect only the relevant files or context I provide. Propose a small safe exercise, wait for my approval before editing, and finish with a summary of what changed, how to test it, and what I should review. Do not touch production, do not commit secrets, and do not make unrelated changes.

Bad Prompt vs Better Prompt vs Expert Prompt

Bad prompt:

Fix this.

Better prompt:

Help me with Model Selection Without Overpromising. Explain what you need to inspect first, then propose a small plan before editing.

Expert prompt:

I want to complete Model Selection Without Overpromising inside this project. Goal: produce a safe, reviewable result for a beginner. Context: I will provide the relevant file, URL, error, or workflow. Constraints: do not edit unrelated files, do not expose secrets, do not deploy, and ask before destructive commands. Done when: you explain the change, list tests to run, identify risks, and give me a rollback note.

Hands-On Exercise

Design a safe expert workflow prompt with one worker, one reviewer, one approval gate, one failure mode, and one evaluation score.

Expected Result

You should have a controlled expert workflow that is explicit about permissions, data boundaries, review, and stopping conditions.

Troubleshooting

• If Codex invents a capability, ask it to cite or verify the current official surface.
• If parallel work creates conflicting edits, pause and split read-only analysis from write tasks.
• If an MCP server or connector fails, check auth, server status, tool permissions, and workspace policy.
• If an eval score improves while quality gets worse, revise the rubric.

Common Mistakes

• Using advanced tooling before the basic workflow is clear.
• Letting multiple agents edit the same files without coordination.
• Giving broad permissions to tools that only need read access.
• Skipping human approval because the workflow feels automated.

Safety Checklist

• Use least privilege for tools, MCP servers, and connectors.
• Keep secrets out of prompts and logs.
• Make destructive actions require explicit approval.
• Prefer read-heavy subagents for parallel work.
• Record assumptions, verification steps, and stop conditions.

Quiz / Checkpoint

Question: What is the safest next step before asking Codex to edit code for model selection without overpromising?

Answer: Give Codex the relevant context, ask it to inspect first, request a short plan, and define how the result will be reviewed and tested.

Navigation

Previous lesson: Approvals, Sandbox Concepts, and Permissions

Next lesson: Non-Interactive and Scripted Workflows

Return to course hub: codex-course-hub.html