System Prompt Test Harness

What it is for

Changing a system prompt without a fixed set of test cases is guessing, and the regression shows up two edits later when you can no longer tell which change caused it. This is the harness: the same cases every time, one edit per run, results written down.

The skill Copy

Use this when tuning a system prompt in AI Studio, before shipping it as an API
call.

Set up:
1. Write down 8-12 inputs you actually expect, including the two or three that
   have gone wrong before. These are your cases; they do not change while you
   are editing the prompt.
2. Save the prompt in Drive after every edit. AI Studio has no diff, so an
   unsaved good version is a lost one.

Loop:
1. Change ONE thing in the system instructions.
2. Run all cases at temperature 0. Anything above 0 makes a regression look
   like sampling noise and you will chase it for an hour.
3. Record, per case: pass, fail, or changed-for-the-worse.
4. If two cases got better and one got worse, you did not improve the prompt —
   you moved the failure. Keep going or revert.

Ship:
- Raise temperature LAST, and only if the task benefits from variety.
- Use **Get code** to export, and check that the exported call carries the
  system instructions in the `system_instruction` field rather than as the
  first user turn. They behave differently.

How to install it Copy

Google AI Studio

1. Open [aistudio.google.com](https://aistudio.google.com) → Create Prompt.
2. Paste the text below into System instructions (not the user turn).
3. Set temperature and output length for the task, then Save —
   the prompt is stored in your Drive and can be reopened.
4. Get code exports the same configuration as an API call.

Did it work for you?

Where it came from