To try an AI coding assistant safely, work in a throwaway folder with nothing private in it, keep the assistant on ask-first settings, build something small, read every change, and finish with a security scan. The whole practice run takes about 30 minutes and teaches you the habits that matter later.
TL;DR: Make a scratch folder. Start the assistant in a read-only or planning mode. Ask for a tiny program. Approve edits one at a time. Scan the result and read the findings. Then delete the folder. You learn how the tool behaves before it ever touches real work.
Definition: throwaway project. A throwaway project is a small practice folder with no real data, no passwords and no connection to your company systems. If it breaks, you delete it and lose nothing.
Why practice on a throwaway project first?
New tools feel safe until the first surprise. An AI coding assistant that can edit files and run commands may do more than you pictured, especially in the first week.
A throwaway project removes the stakes. There is no customer data to leak, no live system to break and no teammate waiting on your work. You can watch what the tool does with permission prompts, and you can see how it handles a mistake.
It also gives you something concrete to show a teammate or a manager. "Here is what happened in a practice run" is more useful than any opinion.
If you are new to the vocabulary, our post on the difference between Claude, ChatGPT, Claude Code and Codex explains what each name means.
What do you need before you start?
Gather these before the clock starts.
- An AI coding assistant you are allowed to use. Check your company's rules first. If you are practicing on your own time, use your own account.
- A terminal or editor. A terminal is a text window where you type commands.
- A new empty folder. Name it something obvious, like
practice-ai. - About 30 minutes. Set a timer. A hard stop keeps you from wandering.
Do not copy in real project files, passwords or customer records. The point is that nothing here matters.
What are the steps, with times?
- Minutes 0 to 3: set the safe settings. Start the assistant inside the practice folder only. Choose the most careful mode available, and do not assume the starting mode is the careful one. Anthropic's documentation says Claude Code's Manual mode asks before most actions that edit files, run shell commands or reach the network, and that its plan mode is suited to exploring before changing anything. OpenAI's documentation says Codex can be switched to read-only mode with the
/permissionscommand when you want to chat or plan without making changes. - Minutes 3 to 8: ask for a plan, not code. Type something like: "I want a small program that reads a list of names from a text file and prints them sorted. Describe your plan. Do not write any code yet." Read the plan. If something looks odd, ask why.
- Minutes 8 to 15: let it write, one approval at a time. Switch to a mode that lets it edit, but keep prompts on. Approve each file edit after reading it. Note every time it asks to run a command, and say no to anything you do not understand.
- Minutes 15 to 18: run the program. Run it yourself. Does it do what you asked? If not, say what is wrong in plain words.
- Minutes 18 to 20: look at the libraries. Open the file that lists the program's dependencies. A dependency is a library of code written by someone else. Count them. Ask yourself if you recognize each name.
- Minutes 20 to 27: scan. Run a security scan on the folder and read the findings. A sample command is shown below.
- Minutes 27 to 30: write three lines. What surprised you? What did the tool ask permission for? What would you change in the settings? Then delete the folder.
How do you run the scan in step 6?
If you want a no-install option, try the free scan at gold.safeguard.sh/scan.
If you prefer the terminal, the Safeguard command line tool installs with one command on macOS and Linux and has a PowerShell installer for Windows. The CLI page shows the macOS and Linux command below. It also suggests you can read the installer script before you run it, which is a good habit with any script you pipe into a shell.
curl -fsSL https://cli.safeguard.sh/install | bash
cd practice-ai
safeguard scan --dir .
The scan builds a list of components, matches them against a database of known flaws, and prints a summary by severity. According to the CLI page, the scan runs on your machine and only the results leave it, never your source code.
Our posts on installing the Safeguard CLI on macOS, Linux and Windows and on your first scan go step by step.
What does a worked example look like?
All numbers here are invented for illustration. Your results will differ.
You ask the assistant for a small program that sorts names. It picks a web framework you did not ask for and adds a package to read text files.
At step 5 you count 14 dependencies, even though your program is 20 lines long. That is normal. Each library you add brings its own libraries. At step 6 the scan finds 2 known flaws, one medium and one low, both in a library you never heard of.
You then do three things:
- Ask the assistant: "Why did you add this library? Can you do this with less?"
- It removes the framework and uses a built-in feature instead. Dependencies drop from 14 to 3.
- You rescan. Zero findings.
Notice what you learned in 30 minutes. AI tools add more than you ask for. Fewer dependencies mean fewer chances for known flaws. And you can ask the tool to do less.
What should you look for during the practice run?
Keep this short list next to you. After the run, answer each question.
| Question | Why it matters |
|---|---|
| Did the tool ask before editing? | Shows whether your settings work |
| Did it ask before running commands? | Commands can change more than files |
| Did it add libraries you did not request? | Each one is new supply chain risk |
| Did it try to reach the internet? | Network access is a decision, not a default |
| Did you understand every change? | If not, the change is too big |
| Did the scan find anything? | Shows what a gate would block |
If you answer "no" or "I do not know" to any row, that row is your homework before using the tool on real work. Our post on read-only plan mode or edit mode for an AI coding agent helps you choose.
Where does Safeguard fit in?
Safeguard checks what ends up in your software. It builds a software bill of materials (a list of every component in an application), scans dependencies for known flaws, and uses reachability analysis to show which flaws your code can actually trigger. When you are ready to go beyond practice, policies and gates can block pull requests that break your rules, and fix pull requests can be turned on. Fixes are off by default, with every category set to Manual.
The Safeguard CLI can also write code, so you can practice the same cycle with it: ask for a change, read the diff, scan, then decide. The download page lists the installers for the CLI, desktop app, editor extensions and mobile apps.
Try it yourself
Open the demo below to see the local runner in the Safeguard CLI. Look for what stays on the machine and what is sent back.
Click through it at your own pace, or open it full screen.
Quick check
Do I need to know how to code to do this exercise?
No. The goal is to watch how the tool behaves and to practice reading what it changes. If a change confuses you, ask the assistant to explain it in plain words.
What if the assistant asks to do something I do not understand?
Say no, then ask it why it wants to. A good rule is that you only approve actions you could explain to a teammate.
Is it safe to use my real project for the first try?
It is better to start with a throwaway folder. Once you know how the tool asks permission and what it changes, you can move to a real project with the same careful settings.
How often should I repeat the exercise?
Repeat it when you change tools, change settings, or when the tool adds new features. Ten minutes of practice can save you from a surprise on real work.
Sources
- Claude Code permission modes, Anthropic documentation
- Codex agent approvals and security, OpenAI developer documentation
- Safeguard CLI
- Safeguard downloads
Last reviewed: 19 October 2026.
