Stay in control: plans, approvals, undo

Newbie to Tech · Step 2 of 8 · about 5 minutes to read, 15 minutes to do Practice project: https://github.com/jazken/mood-habit-tracker · branch: main. This change is practice. You throw it away at the end. Works with Kiro, Codex, Claude Code or Antigravity. Tool names and buttons were checked against each tool's official docs on 20 Oct 2026 (SGT). Screens can change, so if yours looks different, trust the docs.

Do this now (you can read the rest after)

1. Open your tool in the mood-habit-tracker folder. Stay on main.
2. Turn on its "ask me first" setting. The boxes below show where.
3. Ask:

Plan how to add a fourth habit button called "Stretch" to index.html.
It should look and work like the other habit buttons.
Don't change anything yet. Tell me which lines you'd change.

4. Read the plan. Change one thing, for example: "Call it 'Stretch for 5 minutes' instead."
5. Say go. When it asks to change files, read the diff and say yes.
6. Refresh index.html in your browser. Tap your new habit. Does it get a tick?
7. Now undo it (see "Habit 3" below).

Check you did it: after the undo, the page shows only the three first habits again. You can name the three things you did: plan, approve, undo.

---

Why this matters

An AI coding tool is like a very fast helper who never gets tired. That's great, but fast helpers can also make fast mistakes. You stay in charge with three habits:

1. Plan first. See what it will do before it does it.
2. Approve. It asks, and you say yes or no.
3. Undo. If it goes wrong, you go back.

Every tool has these. They just use different names.

Habit 1: Plan first

A plan is a list of what the tool means to do. Reading it takes 1 minute. Fixing the wrong code can take an hour.

What to look for in a plan:

• Does it match what you asked for?
• Is it doing more than you asked? Say "only add the button."
• Does it mention files you didn't expect?

In your tool: getting a plan

• Kiro: start a spec. You get requirements.md (what), design.md (how) and tasks.md (steps). Kiro waits for you to be happy with each one before it moves on.
• Claude Code: switch to Plan mode. It reads and plans, but doesn't edit your files.
• Antigravity: start in Planning mode. You get an implementation plan. Add comments, then click Review to send them, or Proceed to go. In the CLI, type /plan.
• Codex: switch to Plan mode. It reads and plans, but doesn't edit your files. Use /permissions to pick read-only when you don't want any changes.

Habit 2: Approve

Each tool has a setting that decides how often it stops to ask you. While you're learning, pick the setting that asks the most. It's slower, but you'll see everything.

When it asks, you'll usually see one of these:

• A file change. Read the diff (+ added, − removed). Say yes if it matches the plan.
• A command. Read it first. If you don't know it, ask: "What does this command do?"

Two "never while learning" rules:

1. Don't pick "always allow" for commands that delete things, use sudo (admin power), or send work online (git push).
2. Don't turn on "never ask me" or "skip permissions" modes.

In your tool: the "ask me first" setting

• Kiro: turn on Supervised mode (Autopilot is the default). Supervised stops after each turn that edits files, and you can accept or reject each part. Commands ask Allow or Deny, and sometimes also Always allow or Always deny.
• Claude Code: use Manual mode. Newer versions may start in auto mode, so check. In the terminal, press Shift+Tab to change modes. You may also see "Yes, and don't ask again". Skip that one for now.
• Antigravity: set the artifact review policy to Request review, not Always proceed. Permission rules sort actions into Deny (blocked), Ask (it checks with you) and Allow (no question). The Request Review preset always asks before terminal commands. In the CLI, Shift+Tab cycles default, accept-edits and plan. Stay on default, which shows you each file edit.
• Codex: the usual Auto preset edits files and runs commands inside your folder without asking. It asks before working outside the folder or using the internet. For full "look, don't touch", use read-only. Codex can also be set to never ask. Don't use that while learning.

Habit 3: Undo

Some tools have their own undo, but Git is the undo that works everywhere. You'll learn it properly in Step 3. A good habit is to save with Git before you ask the AI for a big change.

Your project is already a Git repo, because you cloned it in Step 0. So everyone can use this undo for today's change. Ask your tool:

Show me what changed in index.html since the last commit.
Then throw away that unsaved change with git restore.
Explain the command before you run it.

⚠️ git restore deletes unsaved changes for good. That's what you want today. Don't use it on work you want to keep.

In your tool: undo

• Kiro: each message you send makes a checkpoint. Click Restore to go back. Kiro doesn't track changes made by commands it ran, or edits you made yourself.
• Claude Code: type /rewind, or press Esc twice when the input box is empty. It tracks edits made by its file tools.
• Antigravity: I couldn't find a checkpoint or undo feature in the docs. Use Git to undo.
• Codex: the docs say to "create Git checkpoints before and after a task so you can revert changes." Use Git.

Bonus: a rules file

Tired of repeating "keep it simple, explain your changes"? Most tools read a rules file in your project, so you can write your wishes once.

• Kiro: steering files in .kiro/steering/
• Claude Code: CLAUDE.md (it can also read AGENTS.md)
• Antigravity: AGENTS.md or GEMINI.md
• Codex: AGENTS.md

Try putting this line in yours: "I'm new to coding. Explain each change in plain words before you make it."

Words from this step

Word	What it means
Plan / spec	A list of what the AI will do, before it does it
Approve	Say yes to one change or command
Mode	A setting for how often the AI asks you first
Checkpoint	A save point your tool makes automatically
Rules file	A note in your project that the AI reads every time

Go deeper

• Kiro: https://kiro.dev/docs/ide/chat/autopilot/ · https://kiro.dev/docs/checkpoints/ · https://kiro.dev/docs/steering/
• Claude Code: https://code.claude.com/docs/en/permission-modes · https://code.claude.com/docs/en/checkpointing
• Antigravity: https://antigravity.google/docs/artifact-review · https://antigravity.google/docs/permissions · https://antigravity.google/docs/cli/modes
• Codex: https://learn.chatgpt.com/docs/agent-approvals-security · https://learn.chatgpt.com/docs/agent-configuration/agents-md

Next: Step 3, Git part 1: save points you control.