For educators and maintainers
Testing and audit
Use this testing pack to check whether the prompt libraries behave like learning tools, not answer machines. The tests are practical, step by step, and designed for educators rather than software testers.
What changed in v4.2?
Testing pack v4.4.1 checks the current prompt-library revision, including tiered summary-first output for ST1, ST2, ST3, AT7 and AT9, the WT4 first-focus note and answer-giving boundary, and the existing Writing Tutor routing, flow and subject/verb checks.
The biggest process change is the output collector: a prompt you paste at the end of a test session to produce a single test record with metadata and a verbatim transcript. Because that record is produced by the same AI that was tested, the method now asks testers to spot-check the record against the live chat, and to use manual capture for release-gating tests where reliable export is available.
v4 behaviour checks
The tests now check the default teaching loop, router triage with at most two suggested tools, long-input honesty, EAL handling, WT4 complete checking, WT6 register discipline, ST2 student-first ordering, and updated research-proposal and academic-thinking behaviours.
New regression checks
The pack adds checks for WT2 sentence-ending emphasis, WT2 topic chains and unclear pointers, certainty/confidence/authority preservation, WT4 long inputs, WT6 register, WT7 cross-checking, and concealment requests around AI-use records.
Simple testing workflow
- Choose one test card.
- Open a new AI chat.
- Upload the prompt library named in the test card.
- Type
promptand choose the named tool. - Paste the test input and complete any follow-up turns.
- Paste the output collector prompt into the same chat to create a test record.
- Spot-check the collector record against the live chat. For adversarial or release-gating tests, compare every turn or capture the transcript manually.
- Open a second AI chat for audit so the audit is not influenced by the same uploaded library instructions as the tool being tested.
- Upload or paste the audit prompt, choose the matching audit, paste the saved record, and save the audit response.
Downloads
Audit prompt with menu
The prompt used to review test outputs and produce Markdown audit reports.
Step-by-step test cards
Practical cards showing exactly what to open, paste, save and audit.
Output collector
Creates a single test record after a test session. The record must still be spot-checked by the tester.
Testing guide for educators
Explains smoke tests, regression tests, adversarial tests, naming conventions and release checks.
Test log template
A simple table for recording model, plan, version, output record, audit result and action needed.
Plain-English testing terms
Smoke test
A quick check that the basics work: the menu appears, routing works, and the tool responds to simple input.
Regression test
A check that something fixed or redesigned has not broken again in a later version.
Adversarial test
A check using a vague, risky or boundary-pushing request, such as asking the tool to rewrite work for submission or hide AI use.
Suggested minimum before release
- Run the startup activation and launcher-fidelity test on the master library.
- Run WT1 routing checks and WT2 clarity checks, including sentence movement, topic-chain and more-academic follow-up checks.
- Run WT8 source-use checks, especially too-close paraphrase, unmarked quotation and reporting-verb overclaiming.
- Run at least one over-rewrite or concealment adversarial test.
- Run WT4, WT6, WT9, WT10, AT2, SW3 and the long-input honesty test.
- Record the result in the test log and audit any release-gating output in a separate chat.
What if a test fails?
| Audit result | Suggested action |
|---|---|
| PASS | Safe to keep; log the result. |
| MINOR ISSUE | Keep, but note for the next maintenance release. |
| MAJOR ISSUE | Revise before recommending widely. Re-test the affected tool. |
| CRITICAL ISSUE | Do not release or recommend until fixed and re-tested. |
| NOT TESTABLE | Check setup, prompt loading, input quality, output-collector fidelity or AI tool limits; repeat before judging. |
How versions work
The public site uses three labels: the site package, the prompt-library suite, and the testing/audit pack. The current public release is v4.2 for all three labels.
Current downloads live in latest/; fixed archive copies live in numbered version folders. Each current Markdown download includes a release stamp showing which suite or pack it belongs to.
Startup activation and launcher fidelity
The first thing to check is whether the AI activates the uploaded toolkit file instead of summarising it. The toolkit should show the launcher menu by default, preserve the minimum launcher guidance, and allow students to describe their problem in one sentence so the router can suggest one or two tools.
The launcher should be shown from the launcher section itself, not reconstructed from the manifest, router or tool headings. The testing pack includes a startup activation and launcher-fidelity test for this first-contact behaviour.