Cases before confidence
Start with a task, not a blank test set.
Each pack includes two prompts, inspectable cases, expected fields and the policy behind them. Preview in the Test Bench, or take portable files to your own runner. Free reuse, including commercial adaptation, under each pack's stated permission.
Meeting notes → action register
Does a proposed task become an invented commitment? Twelve cases cover unassigned actions, corrections, completed work, dependencies and missing deadlines.
- Subset expected values for IDs/owners/deadlines plus required-field/type structure checks.
- Three manual criteria for useful task wording and preserved scope.
- Empty outputs and ambiguous owners stay meaningful.
Support tickets → typed records
Does a refund mention become a refund request? Sixteen cases cover order corrections, current intent, numeric types, missing values and untrusted instructions.
- Exact JSON checks for five typed fields.
- A reason for each task-specific expected record.
- Generic JSON, project JSON, CSV and rendered prompts.
Try a real-source editorial exercise
Effort is not a promised result: six historical Rust compiler-team excerpts, pinned sources and MIT notice, authored first-pass mistakes and reviewed revisions. See how a task passes JSON checks while overstating a commitment, and keep a start date out of the deadline field. Sources are real; responses and judgments are agent-authored, not independent model evidence or current assignments.
Inspect the contract before scoring a response
These are agent-authored development examples, not independently reviewed ground truth or measured model benchmarks. Projects start without model responses or scores. An explicit policy makes disagreements reviewable; it does not make one policy universally right.
- Preview every case and both prompts. Preview does not replace your current or saved project.
- Download current work before confirming starter loading. Loading replaces the tab's project, turns Remember off and leaves the previous saved copy intact.
- Run the prompts yourself with your chosen model/settings. Do not send expected answers or rationales as model input.
- Read responses as well as automatic checks. Save a reviewed project and share an offline evidence handoff.
Adapt the policy and examples to your workflow, then reserve unseen cases for validation. There is no account, API key requirement or model call inside PromptKB. Other services you use have their own privacy and costs.