AI does
Drafting, variation and consistency checking.
You do
Diagnose learning, maintain the relationship and close the feedback loop.
1 Research
Why this matters, and what good looks like
EEF warns that poorly designed feedback can harm progress and consume time unproductively. DfE notes that generative AI may produce inaccurate, unsafe, biased or decontextualized content. Teachers therefore select evidence, diagnose learning and decide how the learner will use feedback.
These receipts support feedback safeguards, not AI feedback accuracy or workload savings. They do not establish how often AI feedback fails. UK guidance is not Belgian school policy. This method drafts new wording from evidence in scratch mode; it does not assume an existing feedback note.
- Start from teacher-selected, verified evidence.
- Link wording to one learning intention or criterion.
- Keep one usable next action.
- Separate strength from the evidence's limit.
- Plan the next opportunity to use feedback.
Where the evidence comes from
Strong for feedback practice; AI benefit limited/early.
2 Workflow
Brief it, steer it, check it
Confirm the evidence boundary
Select the excerpt and diagnose the learning yourself. Never enter personal data, confidential records or identifiable work into AI tools. A bounded agent may draft only your preselected, prediagnosed evidence after approval; it cannot triage work or diagnose learning.
Prompt 1 · Evidence boundaryCONTEXT Learning intention or criterion: [paste exact text]. Teacher-selected non-identifying work excerpt and teacher-verified supporting evidence: [paste exact text]. Relevant task context: [paste it]. Teacher diagnosis, uncertainty and need for direct interaction: [paste decisions]. Teacher permission/privacy confirmation and approval to draft this case: [paste decisions or unknown]. REQUEST Separate quoted verified evidence, attributed teacher interpretation and unknowns. Check whether the supplied teacher decisions permit drafting. HOLD if evidence selection, diagnosis, privacy, permission or drafting approval is missing, or direct interaction is required first. QUALITY BAR Never select new evidence, triage work or diagnose learning. Do not infer effort, intent, ability, motivation or authorship. Add no evidence or grades. Missing information stays unknown. Approval here permits drafting only, never release. FORMAT Three blocks: verified evidence with exact work/source quotes; teacher interpretation with exact teacher quotes; unknowns or needs conversation. Then separate gate lines for teacher selection, diagnosis, privacy, permission, drafting approval and direct interaction; each needs a source quote and HOLD or draft-only status.
Draft focused feedback
Use only the checked boundary after teacher approval. Choose statement or question wording for the same single next action. A written note and audio script are alternatives, not two tasks for the learner.
Prompt 2 · Feedback draftCONTEXT Checked evidence boundary and teacher drafting approval: [paste full record]. Exact work/source text and teacher diagnosis: [paste them]. Learning intention or criterion: [paste it]. Previous feedback: [paste it or none]. Next opportunity, time and support: [paste details]. Teacher-selected single next action: [paste one action]. Teacher-selected wording option, tone and access requirements: [statement or question; add requirements]. REQUEST Draft one evidence-linked strength, one boundary and exactly one actionable next step. Use the teacher's selected statement or question wording. Produce written and audio alternatives with the same content. QUALITY BAR HOLD for missing approval, unsupported diagnosis, a required conversation, an infeasible action or multiple actions. Do not split a bundled task into hidden extra verbs or add substeps. Ask the teacher to narrow it. Use only supplied evidence, context and teacher choices. Add no grade, trait, prediction or diagnosis. FORMAT Written note and audio script, each three labelled lines: strength; boundary; next step. Then provenance table: exact output sentence | exact work/source or teacher-decision quote | learning-intention link. Mark teacher approval required before release.
Check accuracy and load
Check quotations and the action count in both versions. Unsupported, inaccurate, overwhelming or relationship-replacing feedback blocks release. Retain the original draft; record teacher edits separately and rerun the audit before approval.
Prompt 3 · Feedback auditCONTEXT Exact work/source text, teacher diagnosis and checked boundary: [paste all]. Draft written note and audio script: [paste exact drafts]. Learning intention: [paste it]. Teacher-selected next action, opportunity, time, support and interaction decision: [paste decisions]. REQUEST Audit every sentence for support, accurate quotation, specificity and feasibility. Count independently required student actions in each version. Flag bundled actions, unsupported claims and any need for direct teacher interaction. QUALITY BAR Quote both sides exactly, including verified lines. If no supporting quote exists write absent; never paraphrase a quotation check. Do not rewrite feedback, invent support or approve release. Block inaccurate, unsupported or overwhelming wording and required interactions that drafting would replace. FORMAT Release blockers first, then optional edits, then verified lines. For every line: exact output quote | exact source/teacher quote or absent | check | finding. End with each version's action count and action quote, plus separate accuracy, support, load, feasibility and interaction decisions. Teacher approval remains required.
Close the loop
Resolve blockers and approve the final note yourself. Decide how the student will use it. Keep necessary teaching or conversation with you. No agent sends feedback automatically or treats a passed audit as teacher approval.
Prompt 4 · Use planCONTEXT Teacher-approved feedback and explicit approval record: [paste exact text]. Resolved audit and any teacher edits: [paste record]. Approved single next action and next opportunity: [paste them]. Time, support and expected evidence of that same action: [paste details]. REQUEST Draft a brief use routine for the approved action only. HOLD for absent approval or unresolved blockers. QUALITY BAR No new student action, reflection, prediction or assessment requirement. The evidence line names the product of the same action, not a further task. The teacher checks use and handles any needed conversation. Do not send feedback or make consequential judgements. FORMAT Three lines: student action; evidence of that action; teacher follow-up. Then approval provenance quote and teacher-only release reminder.
Check before you use it
- Did the teacher select evidence, diagnose learning and approve drafting?
- Does every sentence have an exact source or teacher-decision quote?
- Is there exactly one feasible next action in each alternative?
- Are questions available without adding another learner task?
- Are all audit blockers resolved before explicit teacher release approval?
- Does the teacher retain required conversation and follow-up?
Stop rule
Stop if feedback is factually wrong, unsupported, overwhelming or substitutes for direct teacher interaction.
3 Reuse it
So the next one takes minutes
Build a skill
Save this as a reusable instruction you can run on any project.
A teacher can trigger the same evidence-first drafting method for one work sample at a time.
Build an agent
Chain the steps, with a checkpoint where you approve.
Only drafts from teacher-selected, prediagnosed evidence with explicit drafting approval; never selects or triages work, diagnoses learning or sends feedback. Teacher approval is required before release.
Use cases are starting points with one perspective. Disagree with the process, that's expected. Make it yours.
Co-funded by the European Union under Erasmus+ KA210-SCH. No student data. No teacher material stored.