Home / Free tools / Instruction test bench
Free · No signup · Browser only

Instruction test bench

See whether your custom instructions actually improve answers. Run the same prompt before and after, paste both replies, and compare them with hidden A/B 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.

Review or copy the JSON report

Review your output

Your output will appear here.

How to use the instruction test bench

  1. Pick a starter probe (or write your own task) and run it in two fresh chats: one without your custom rules and one with them.
  2. Paste both replies below and click Compare with hidden labels.
  3. Pick the better answer (or a tie) to reveal which reply used the instructions, along with word and bullet counts. Export your trials as JSON if desired.

A concrete example

Test a “3 bullets max” rule on a rambling status update, then verify that the shorter reply still keeps every key risk.

Questions and limits

Does AgentTune call a model or charge for comparisons?

No. You run the prompts in your own AI account and paste the replies here. Everything runs locally in your browser.

What can one comparison prove?

One pair of replies is a quick sanity check. Run 5 to 10 tasks across your real workflow to see whether an instruction consistently helps without hurting accuracy.

What is included in the report?

The local JSON download includes your task prompt, evaluation criteria, tested instructions, both replies, word counts, and your A/B votes.

Go further

A protocol for testing Claude preferences explains the method and practical limits.

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.