benchrStart from real work
The open packs provide fixed synthetic cases and scoring rubrics. They are starting material—not proof that one prompt or model wins.
Prompt contract · A/B discipline · browser-only
Build a prompt you can actually test.
Turn an idea into an explicit task, output contract, evidence boundary, and success criteria. Capture version A, make a controlled change, then send one or both versions to Labs.
Preparing the workbench…
01 · Define
Prompt contract
Start with an open evaluation case or write your own.
Change one variable
Capture the exact baseline, change one section, and keep the model, settings, cases, and scoring rubric fixed. If several sections move, treat the result as exploratory.
Measure the failure
Score observable criteria: parse failures, missing fields, unsupported claims, policy breaches, and reviewer acceptance. Do not substitute a completeness checklist for an evaluation.
Why the builder is structured this way
Current provider guidance differs by model, but repeatedly emphasizes clear instructions, explicit constraints and output shape, relevant context, and empirical testing. OpenAI also recommends outcome-first prompts and rerunning evals as prompt changes are isolated; Anthropic starts prompt engineering with measurable success criteria and tests; Google describes prompting as iterative and recommends breaking complex prompts into components.
No pattern guarantees an improvement. Model behavior, tool definitions, generation settings, and the task itself all matter, so the Workbench produces a testable artifact rather than a universal “better prompt” claim.
OpenAI model prompting guidanceAnthropic prompting best practicesAnthropic evaluation criteriaGoogle prompt design strategies
Source guidance reviewed August 26, 2026. Always recheck the documentation for the exact model and API you deploy.