Skip to content

Guide / Prompt comparison

Compare AI agent prompts in one fixed environment.

A prompt comparison is interpretable only when the model, wrapper, tools, environment, and run protocol remain fixed. This guide isolates the prompt as the named variable and preserves the trajectory evidence behind each outcome.

For
Agent developers testing a prompt revision against a current baseline
Outcome
A paired prompt comparison with explicit controls and inspectable runs

1 / Variable

Describe the prompt change in operational terms

State what instruction changed and which observable behavior it is expected to affect without predicting the result.

  • Keep the model and model settings unchanged.
  • Keep tool descriptions, wrapper logic, and retry policy unchanged.
  • Version both prompt texts or store stable digests.
  • Name one primary behavior before running either condition.

2 / Controls

Pair prompt conditions against the same environments

Use the same ordered set of game versions, configurations, deterministic inputs, and stopping rules for the baseline and candidate prompt.

The environment can be reproduced, but the agent may still choose different actions. Once choices diverge, later observations can diverge too; record the branch instead of treating later turns as identical inputs.

3 / Evidence

Compare action behavior before comparing endings

  • Legal and rejected action rates
  • Tool reads before consequential writes
  • Response to stale versions and typed errors
  • Policy consistency across comparable observations
  • Ending point movement, reported separately from trajectory behavior

4 / Report

Report the prompt effect without widening the claim

Describe the observed difference under the named WagerCall versions, run protocol, and sample rather than declaring one prompt universally better.

  • Publish both prompt identities and all environment controls.
  • Include attempted, rejected, incomplete, and completed runs.
  • Name any post-run exclusions and why they were necessary.
  • Keep derived measures separate from WagerCall's objective record.

Next step

Run the paired comparison

Apply the same fixed protocol to both prompt versions.

Run the paired comparison