Product brief · November 3, 2026
Six ways to develop and review your prompts.
Selected changes for people who want repeatable cases, usable outputs and reviewable evidence—not just a prompt that worked once.
Covers six selected updates from the public record through 19:17 UTC, not every deployment or a full week. Written and self-reviewed by the agent using the Test Bench. These are product capabilities, not measured human benefits or a model benchmark. Inspect the editorial review and its source notes.
Test Bench
Change: The Test Bench keeps two prompt variants, reusable cases and recorded responses together.
Why it matters: You can inspect a repeated task across cases instead of relying on one impressive answer.
Next: Open the Test Bench at /lab/ and choose New blank project; add an ordinary case, a difficult case and a missing-evidence case.
Limit: Run prompts in your own model and paste the responses. The bench does not execute or verify model runs.
Typed extraction checks
Change: JSON checks compare typed expected values and explain missing, invented or mistyped fields.
Why it matters: A response containing the right words may still be unusable as a structured record.
Next: Open /lab/extraction/ and download the support-ticket starter, then inspect its expected values before collecting responses.
Limit: Passing means matching your configured contract, not proving that the contract or answer is factually correct.
Spreadsheet case sets
Change: CSV and spreadsheet TSV imports build a case set through column mapping and preview.
Why it matters: You can prepare cases as a table rather than enter each input separately.
Next: In the Test Bench, expand Import a case dataset from CSV or a spreadsheet, read the columns and preview before choosing append or replace.
Limit: Matching names are not merged, and new cases start without recorded responses or reviews.
Support extraction pack
Change: The support-extraction pack provides sixteen fictional cases, typed expectations, two prompts and policy rationales.
Why it matters: Its portable downloads provide a task-specific starting set inside or outside PromptKB, with free commercial adaptation.
Next: Open /packs/support-extraction/, download the project and check the refund, issue-priority and amount policies against your own task before using its expectations.
Limit: The policies are authored choices, not universal ground truth or measured model accuracy.
Saved-baseline revision review
Change: Saved-baseline comparison puts a previous project beside the current candidate without replacing either.
Why it matters: You can inspect regressions and missing evidence across the case set before adopting a revision.
Next: Keep a project download before editing; later open Compare a revision against a saved baseline below the matrix and load that earlier file.
Limit: Only exact unique case-name/input matches with current recordings qualify, and both outputs use today's candidate checks—not historical pass rates.
Context-bound manual criteria
Change: Manual criteria retain the basis for a judgment when the review policy changes.
Why it matters: Earlier assessments and notes remain inspectable instead of being silently treated as current.
Next: Expand Define shared manual review criteria, apply your criteria, then start a review of a recorded response and assess every requirement before deciding.
Limit: These are reviewer judgments, not automated semantic verification. Changing criteria or case checks can make a bound review out of date without a new model run.