Analyze student, family and partner feedback

1

Research

Why this matters, and what good looks like

Use one existing feedback round, one question and up to four safe pooled evidence rows. A pooled row combines responses without individual or subgroup identifiers. Bring private originals and sampling notes. If essential checks exceed 60 active minutes, stop and narrow the question. New consultation needs a separate plan; never replace missing voices with AI guesses.

These sources offer professional guidance. PBLWorks discusses process learning; EEF addresses teacher-to-pupil feedback, not stakeholder survey analysis. England's DfE guidance requires local data-protection judgement. None validates this coding method, an anonymization threshold or AI benefit. Satisfaction is not evidence of learning.

PBLWorks cautions against focusing on the final product alone. Keep process questions visible.
EEF notes feedback can support progress or harm it and consume time. This does not validate stakeholder coding.
DfE calls for transparency and local data protection officer or IT guidance. Consent alone cannot establish safe processing.

Where the evidence comes from

Moderate-strong for stakeholder voice; AI benefit unvalidated.

2

Workflow

Brief it, steer it, check it

Minimize before using AI

Privately check consent, local permission and identification risks first. Remove names, rare combinations and small cells. Use optional L06, L07 and L08 feedback-route records or equivalent documents; keep contact lists outside AI.

Prompt 1 · Inventory

CONTEXT
Teacher-minimized pooled feedback with safe source IDs, exact safe quotes, counts and denominators: [records].
Teacher privacy and provenance gate, sampling limits and missing voices: [gate].
REQUEST
Organize only the approved safe evidence and its stated limits.
QUALITY BAR
If the human gate is missing, provenance unclear or any identification risk remains, return HOLD without repeating sensitive text. Do not anonymize raw personal data here. Consent alone is insufficient. Never invent quotes, counts, missing voices or a privacy threshold. Copy counts with their units and denominators; never infer people from mentions.
FORMAT
Table: Source ID | Exact safe quote | Count and denominator | Sampling limit. Then Missing voices. For blocked input return HOLD and required human check only.

Propose themes without losing dissent

Privately verify every quote and count against originals. Supply the checked inventory and corrections. Decide whether evidence can support a theme; ask for disconfirming evidence, meaning feedback that challenges the proposed theme.

Prompt 2 · Themes

CONTEXT
Original approved safe feedback: [records].
Exact approved step 1 output: [previous].
Teacher verification, corrections and adequacy decisions for these source rows: [decision].
REQUEST
Propose at most two themes and show dissent, missing voices and limitations for each.
QUALITY BAR
Return HOLD if verification or adequacy decisions are absent or evidence is inadequate. Keep source IDs, quotes, counts, units and denominators exact. No subgroup reconstruction or counts inferred from quotes. Themes are proposals; satisfaction cannot establish learning. Do not decide action or claim representative consensus.
FORMAT
Table: Theme ID | Proposed theme | Source IDs | Exact quote and count | Dissent | Missing voices | Limit.

Record your response decisions

Choose themes and responses yourself. Check whether absent voices require further consultation before action. Supply your choices, implications and unresolved questions. Do not let AI decide what feedback merits action.

Prompt 3 · Matrix

CONTEXT
Original approved safe feedback: [records].
Exact approved step 2 output: [previous].
Teacher-selected themes, edits, implications, response decisions and unresolved questions: [decision].
REQUEST
Format the supplied choices as the stakeholder feedback matrix. Preserve supporting and disconfirming evidence.
QUALITY BAR
Missing decisions mean HOLD; do not choose a response. Preserve quotes and count units exactly. Do not attribute opinions to missing groups, imply consensus, infer learning or create commitments. Keep unsafe details out. Mark unverified interpretations as such and retain unresolved questions.
FORMAT
Table: Theme | Supporting quote/count and source ID | Dissent | Implication | Response | Unresolved question. Then Sampling and missing-voice limits.

Check your selected matrix

Select and edit the matrix. Privately recheck sampling, exact quotes, counts and privacy before this comparison. Resolve flags yourself. Decide whether further stakeholder checking is needed before any use or distribution.

Prompt 4 · Audit

CONTEXT
Original approved safe feedback: [records].
Exact step 3 output reviewed by the teacher: [previous].
Teacher-selected matrix, edits and private quote/count/sampling/privacy verification: [decision].
REQUEST
Compare the selected matrix with sources and decisions. Check exact quotes, counts, units, dissent, missing voices and unsupported learning or consensus claims.
QUALITY BAR
Return HOLD if selection or human verification is missing. Do not rewrite, choose actions, certify anonymity or approve distribution. Flag unsafe content without reproducing it. A textual match is not independent source verification. No inference about identifiable people or groups.
FORMAT
If selection or human verification is missing, output only HOLD, the missing input and the required human check without repeating sensitive text. Otherwise output, in order: a table with exactly these columns: Check | Source ID or decision | Finding | Human action; a Remaining limits section; one conclusion sentence; and the final line Human release pending. Finding must be Text match, Mismatch or Unresolved. Text match means agreement with supplied text, never verified fact. Always retain the table and limits, including when every finding is Text match. In that case the conclusion must be exactly: The supplied-text comparison found no mismatch. Otherwise it must be exactly: The supplied-text comparison found mismatches or unresolved checks. Put provenance limits, including any synthetic-input status, in Remaining limits. Do not add a preamble or other sections.

Check before you use it

Did you minimize data privately before using AI, including small cells and rare combinations?
Have you checked every quote, count, unit and denominator against originals?
Are sampling limits, missing voices and dissent still visible?
Have you kept satisfaction separate from learning evidence?
Are implications qualified and responses your explicit decisions?
Have you resolved privacy and provenance concerns before retaining or sharing the matrix?

Stop rule

Stop when provenance is unclear, small groups risk identification or a theme rests on too little evidence.

3

Reuse it

So the next one takes minutes

Suitable

Build a skill

Save this as a reusable instruction you can run on any project.

The checking sequence can be reused with fresh safe inputs and explicit teacher decisions.

Not suitable

Build an agent

Chain the steps, with a checkpoint where you approve.

Private source checks and interpretation require direct human control within this bounded task.