Prioritize redesign changes and next-run tests

Where the evidence comes from

• Instructional improvement cycle: A teacher’s toolkit for collecting and analyzing data on instructional strategies
• Implementation and process evaluation (IPE) for interventions ineducation settings: An introductory handbook

Strong for inquiry/improvement cycles; AI benefit unvalidated.

1. Fix the redesign boundary

   Bring up to three verified problem notes and one project segment. Optional I01 findings and I03 maps help; equivalent teacher evidence notes work. Privately remove identifying information. Name essential project-based learning features, access routes and a workload ceiling.

   Prompt 1 · Options Copy

   CONTEXT
   My safe, verified evidence notes with IDs, original-check record and limitations: [evidence notes]
   My project segment, essential learning and PBL features, access routes, workload ceiling and explicit safe-input decision: [constraints and gate]
   REQUEST
   Propose at most three change options linked to observed problems and evidence IDs. Explain trade-offs and assumptions. Do not rank or select options. Keep a no-change option available if the evidence is insufficient.
   QUALITY BAR
   Stop with HOLD if evidence, human original checks or constraints are missing or unsafe. Never invent results, approvals, learner needs, resources or commitments. Do not grade or infer causality. Reject proposals lacking evidence, removing essential PBL features or creating workload/access harms. All estimates and options are proposals for human checking.
   FORMAT
   If blocked, only HOLD | Missing or unsafe input | Teacher next action. Otherwise a table: Option ID | Problem | Evidence IDs and limit | Proposed change | Protected features and access | Workload estimate | Trade-off or assumption. End with Teacher selection needed.

2. Record your priorities

   Check evidence and feasibility yourself. Rank the backlog, selecting one or two changes and recording why others wait. Recalculate time and resource totals independently. Keep student inquiry, choice, critique and revision; confirm access with people outside AI where needed.

   Prompt 2 · Backlog Copy

   CONTEXT
   Exact step 1 output: [options output]
   My explicit corrected ranking, one or two selected changes, rejected/deferred options, evidence checks, access/PBL checks and independent arithmetic: [priority decision]
   REQUEST
   Format my ranked backlog without changing rank or choosing for me. Retain evidence links, reasons and constraints. Mark each item selected, deferred or rejected exactly as I decided. Surface any contradiction in my decision.
   QUALITY BAR
   If the decision is missing, selects more than two changes, lacks independent arithmetic or fails evidence/PBL/access/workload checks, stop with HOLD. You cannot verify human calculations or commitments. Do not invent totals, prioritise on my behalf or remove protected features. An unverified estimate remains HOLD.
   FORMAT
   If blocked, only HOLD | Missing or unsafe input | Teacher next action. Otherwise a table: Rank | Option ID | Evidence IDs | Teacher decision | Teacher reason | Verified constraints and arithmetic | Remaining limit. End with Teacher test design needed.

3. Define prospective test cards

   Choose the prediction, measure, collection point, owner and stop/adapt threshold for each selected change. State how evidence will be checked privately, including access and workload. Keep learning measures separate from activity. These are future tests, not reported results.

   Prompt 3 · Test cards Copy

   CONTEXT
   Exact step 2 output: [backlog output]
   My selected-change confirmation and explicit test decisions, including prediction, measure definition, comparison limits, collection point, owner, human arithmetic and stop/adapt rules: [test decision]
   REQUEST
   Organise one test card per teacher-selected change. Link problem, evidence and proposed adaptation. Preserve my prediction and measure exactly. Name what evidence would challenge the prediction. Preserve essential learning, PBL and access constraints from the backlog.
   QUALITY BAR
   Stop with HOLD if required test decisions or verified constraints are absent. Do not choose thresholds, owners, dates or tests for me. Mark every prediction prospective and every outcome not yet collected. No invented trial results, AI grading or causal claims. Alternative explanations and comparison limits remain visible. Stop/adapt rules must include access and workload harms.
   FORMAT
   If blocked, only HOLD | Missing or unsafe input | Teacher next action. Otherwise a table: Option ID | Problem and evidence | Change | Prospective prediction | Measure | Collection point and owner | Challenging evidence and limits | Stop/adapt rule | Outcome status. End with Human trial approval pending.

4. Check before any trial

   Select the final backlog and cards. Recheck sources, arithmetic, access and feasible ownership privately. Ask for an audit, then decide whether to authorise a trial. Stop and re-scope if planning and necessary human checks exceed 60 active minutes.

   Prompt 4 · Audit Copy

   CONTEXT
   Exact step 3 output: [cards output]
   My selected backlog and cards, explicit edits and final source/PBL/access/workload/owner checks with independent arithmetic: [final decision]
   REQUEST
   Audit my selected plan against the supplied decisions. Flag unsupported changes, lost constraints, totals needing human checks, predictions presented as results, missing measures/owners and weak stop/adapt rules. Do not rewrite, rank or authorise the plan.
   QUALITY BAR
   If selection or any required human check is missing, stop with HOLD. Each finding needs an option ID and a teacher action. No AI grading, inferred learner characteristics or causal claims. Do not treat supplied approvals as a real trial or certify calculations yourself. Even no textual defects leaves trial authorisation with humans.
   FORMAT
   If blocked, only HOLD | Missing or unsafe input | Teacher next action. Otherwise a table: Check | Option ID | Finding | Teacher action. End with Human trial approval pending; no results collected.

Check before you use it

• I selected only one or two changes, each linked to evidence I opened and checked.
• Essential learning, inquiry, student choice, critique, revision and access remain intact.
• I independently checked arithmetic, workload and feasible ownership with the relevant people.
• Predictions, measures and collection points are explicit; no future outcome is reported as a result.
• Each card includes evidence that would challenge it and access/workload stop or adapt rules.
• I retain trial approval and later interpretation; AI neither grades nor establishes causes.