# Check the output before you use it

Web page: https://academy.memberpress.ai/courses/same-skills-any-assistant/l6/

Lesson 6 of 6 in [Same Skills, Any Assistant](https://academy.memberpress.ai/courses/same-skills-any-assistant/), 8 min.

Every assistant on this page can be confidently wrong. The habit that saves you is the same everywhere: treat the answer as a draft until you've checked it.

**The habit**

- For facts: ask where each claim came from, and open the source.
- For writing: read it out loud once. If you wouldn't say it, change it.
- For code: read the diff before you accept it, and run the tests.
- For anything that acts on your behalf: say "plan first, then wait for me".

**Side by side**

- **Claude:** in Claude Code, read the diff before you accept it. The Claude Code course has a lesson on it.
- **ChatGPT and Codex:** Codex's approval setting can ask before it acts (`on-request`), and its sandbox can be read-only.
- **Grok:** ask it to list its sources, then open them yourself.
- **Cursor:** Plan mode drafts a plan for you to review before you click Build; Ask mode never edits.

Checked against each vendor's docs on 2026-10-14: [Codex sandboxing and approvals](https://learn.chatgpt.com/docs/sandboxing), [Cursor Plan mode](https://cursor.com/help/ai-features/plan-mode), [Cursor Ask mode](https://cursor.com/help/ai-features/ask-mode).

## Try these prompts

- Review your last answer as a skeptical editor. List every claim I should check before publishing, and how I'd check each one.
- Before you change anything, write the plan and wait for me to say go.