Build an evidence-backed impact brief

Where the evidence comes from

• EEF implementation and process evaluation handbook
• IES Instructional Improvement Cycle
• Project Zero Looking at Student Work

Strong for implementation reporting and professional learning; AI benefit unvalidated.

1. Bound the evidence and permissions

   Use one run, one audience and at most eight claims. Optional predecessor methods supply curriculum, evidence and next-test records; equivalent teacher documents work. Privately remove identifiable material before AI use, including revealing aggregates. Keep originals outside AI.

   Prompt 1 · Inventory Copy

   CONTEXT
   Safe evidence pack with source IDs, exact excerpts, counts, dates or period, limitations and teacher verification status: [paste safe evidence pack].
   Teacher scope, audience, purpose, privacy check and permission boundaries: [paste scope and permissions record].
   Existing brief: [paste safe existing brief or none].
   REQUEST
   Prepare a candidate claim ledger for this bounded brief. Cover need, project model, reach, implementation, learning evidence, voices, limitations, workload/cost and adaptations. Identify missing fields rather than filling them.
   QUALITY BAR
   Use only supplied evidence; do not verify your own sources. Retain negative findings. Keep anecdotes separate from outcome evidence; infer no causality or universal effectiveness. Preserve exact quotations and numerator/denominator pairs, units and periods. Never invent testimonials, permissions, costs or learner details. If scope, privacy clearance or verification is missing, return HOLD with missing inputs only. An absent optional voice can be recorded as absent without inventing testimony. Mark individual unsupported claims HOLD; do not draft them.
   FORMAT
   If blocked: HOLD | missing input | human action.
   Otherwise use columns: claim ID | proposed statement | source ID and exact excerpt | numerator/denominator or units | limitation | teacher verification | use status.
   Then list missing brief fields and permission boundaries.

2. Check every proposed claim

   Open originals yourself, check arithmetic and quote fidelity, and record explicit corrections or exclusions. Keep a separate permission decision. If core evidence cannot be checked within the bounded session, stop and re-scope rather than compressing verification.

   Prompt 2 · Reconcile Copy

   CONTEXT
   Exact step 1 ledger selected by the teacher: [paste selected inventory output].
   Explicit teacher claim checks, corrections, exclusions and permission decisions after opening originals: [paste versioned teacher checks].
   REQUEST
   Produce the reconciled ledger using only supplied human decisions. Identify any unresolved contradiction or stale verification. Do not choose which disputed account is true.
   QUALITY BAR
   No model-generated verification or permission. Keep each claim ID and its exact supporting excerpt. Show corrected numerator, denominator, unit and period; do not silently change numbers. Exclude rejected claims, retaining reasons. Distinguish descriptive learning evidence, implementation and missing voices. If teacher checks are absent, return HOLD only. Unsupported items stay HOLD. No extrapolation from a local run or anecdote to an effect. Checked wording must be no stronger than its excerpt and supplied teacher decision. Do not add causal verbs or reasons such as led to, because or caused without supplied causal justification; after stays temporal or HOLD. For each changed claim, show before and after wording, source support and the supplied edit decision; flag unresolved strengthening instead of silently repairing it.
   FORMAT
   If blocked: HOLD | missing input | human action.
   Otherwise use columns: claim ID | checked wording | source ID and exact excerpt | numerator/denominator or units | limitation | supplied human decision | use status.
   Finish with excluded claims and unresolved permissions.

3. Draft the teacher's chosen account

   Choose the narrative and bounded next decision yourself. Supply the exact checked ledger and your edits. Draft a short account for the stated audience, retaining evidence gaps, negative findings and conditions another setting would need to examine.

   Prompt 3 · Draft Copy

   CONTEXT
   Exact step 2 ledger selected by the teacher: [paste selected reconciled output].
   Teacher narrative choice, next decision, transfer conditions, authorship wording and explicit edits: [paste versioned narrative decisions].
   REQUEST
   Draft an impact-and-learning brief of at most 350 words for the supplied audience. Use the supplied narrative and next decision without expanding commitments. Adapt an existing brief only where checked evidence supports it.
   QUALITY BAR
   Use only USE claims and cite their IDs beside each factual sentence. Retain exact approved quotes, counts with denominators, periods, workload/cost units, uncertainty and negative findings. Missing voices stay missing. State that one local run does not establish causality or universal effectiveness. Do not add testimonials, learner information, permissions, savings or predictions as observed results. If narrative or next decision is missing, return HOLD only. Compare checked wording with its excerpt and supplied edits: claim strength must not increase. Explicitly reconcile any prior causal overclaim with the teacher-selected temporal correction; without supplied correction or causal justification, HOLD the claim rather than silently substituting a safer quote. This is a draft; the teacher owns the account and distribution.
   FORMAT
   If blocked: HOLD | missing input | human action.
   Otherwise use these labels: Need; Project model and reach; Implementation and adaptations; Learning evidence and voices; Limitations and transfer conditions; Workload/cost; Next decision; Authorship and distribution.
   Finish with columns: sentence or section | claim IDs or supplied decision ID | unresolved issue.

4. Audit the selected brief before release

   Select and edit the draft yourself. Check every claim and permission outside AI, then request a final discrepancy audit. You approve any distribution separately. If active work exceeds the bounded estimate, stop, retain HOLD items and reduce scope.

   Prompt 4 · Audit Copy

   CONTEXT
   Exact step 3 brief selected by the teacher: [paste selected draft output].
   Exact approved step 2 ledger: [paste approved reconciled output].
   Teacher edits and final claim, quote, arithmetic, privacy and permission checks: [paste versioned final human checks].
   REQUEST
   Audit the teacher-selected brief against the supplied ledger and decisions. Identify unsupported sentences, omissions and changed quotes or numbers. Do not rewrite the brief or grant release approval.
   QUALITY BAR
   Check need, model, reach, implementation, learning, voices, negatives, uncertainty, workload/cost, adaptations, transfer conditions, authorship and next decision. Check ledger wording against its own excerpt as well as the brief; a safer brief does not resolve a stronger ledger. Report before/after claim strength and explicit edit authority; causal verbs without supplied causal justification mean HOLD. Separate evidence agreement from real permission to distribute. Missing checks, unsupported impact, learner identification or universal claims mean HOLD; name the human action. Even when supplied checks agree, leave publication to the human and do not send anything.
   FORMAT
   Use columns: check | evidence or decision ID | discrepancy | disposition | human action.
   End with Overall: HOLD or NO DISCREPANCY FOUND IN SUPPLIED MATERIAL. Then state: Human distribution decision remains separate; nothing sent or published.

Check before you use it

• I opened the originals and checked every claim, quotation, count, denominator, unit and period.
• I retained negative findings, missing voices and uncertainty without claiming causal effects or universal effectiveness.
• I checked workload, costs, adaptations and transfer conditions against the supplied records.
• I removed identifying details and revealing aggregates before using AI.
• I chose the narrative and bounded next decision; AI did not approve either.
• I checked permissions and authorship separately; nothing is distributed without human approval.