Why does Copilot keep getting PowerShell scripts wrong?

Access to documentation doesn't guarantee that a model will use it correctly, and the assumption that it was trained on all of Microsoft's PowerShell material isn't something to take for granted. Generated code can look plausible while mixing up syntax, parameters, or commands.

PowerShell versions and modules matter too. Include your PowerShell version, operating system, relevant module versions, and the exact error. Ask it to check commands against the official documentation rather than treating the first answer as authoritative.

Copilot is a product interface, not one fixed model. The available model and settings can differ between Copilot products and configurations, so check what you're using before comparing results.

Even with the same underlying model, different interfaces can provide different context or tools. Comparing the same prompt across setups may be useful, but a difference in output doesn't prove exactly what caused it.

Try asking for a small snippet first instead of a complete script. Have it explain each command and its parameters, then add the next piece once the first one checks out. That makes a wrong assumption much easier to spot than it is in a long generated script.

Keep the requirements specific. Tell it what input you have, what output you want, and whether the script is allowed to change anything. Ask for a preview of changes where possible, and test in a safe environment before running it against real systems.

For the chat history issue, save the version that worked in a script file rather than relying on a later response to reproduce it. If the interface has a copy button, try that too, then compare the pasted text with the original. If the script was regenerated, check what changed before running it again.

Don't assume a longer script is a better one. Ask for the simplest solution and challenge anything that looks unnecessary. If you get an error, give it the exact message and the relevant code, but still check its proposed fix against the command documentation.

And yes, some people get better results with other models or shorter requests. Others still run into bad scripts with those, so switching tools is worth testing, not a guarantee.