Instruction test bench
Do your instructions change the responses you care about? Run the same task with and without them, paste both answers, and compare with randomized labels.
Your input stays in this tab. No model calls or automatic saving. Export anything you want to keep before leaving.
Set up a comparison
Included in your downloaded report.
Response A
Response B
Which response better meets your criterion? Choose before revealing the labels.
Review or copy the JSON report
Review your output
Your output will appear here.
How to use the instruction test bench
- Choose a probe and define what a good answer should do. Use the same model, settings, tools, and task in fresh conversations; change only the instructions you are testing.
- Paste the baseline and modified answers. The tool randomly labels them A and B while you compare. You may recognize your own pasted text, so this is a lightweight review, not full experimental blinding.
- Choose the better response or a tie. Reveal labels and counts. Repeat across tasks and runs; export your observations before leaving the page.
A concrete example
Ask for exactly three bullets. Compare useful content as well as structure: a shorter answer can still be wrong or omit a crucial point.
Questions and limits
Does AgentTune call a model or charge for comparisons?
No. This free tool measures answers you paste. You run the prompts in the AI product you already use, subject to that product’s limits and pricing.
What can one comparison prove?
Very little about a model in general. Word and bullet counts describe the text; your preference records your judgment. Neither proves factual accuracy or a causal effect. Repeat comparable runs and report how you tested.
What is included in the report?
The JSON export contains your task, criteria, model notes, tested instructions, both responses, randomized label order, preference and text counts. It can include private information, so review before sharing.
Go further
A protocol for testing Claude preferences explains the method and practical limits.
- Communication preferences builder — Create instructions without taking a personality test.
- Instruction checker — Find duplicate rules and wording worth revising.
- Muse Soul.md editor — Safely add communication preferences to Muse Soul.md.
For developers and agents: tool catalog (JSON), local API documentation, and reusable JavaScript functions. These process text; reading a file does not authorize installing instructions.