Build a transparent rubric that explains judgment criteria rather than delegating grades to a model.
This guide proposes a workflow to make assessment expectations visible to learners and assessors. It is a planning pattern, not a claimed firsthand test, guarantee, or substitute for professional judgment.
You need learning outcomes, institutional policy, and assessment owner. Do not paste passwords, API keys, health information, government identifiers, or unredacted customer/employee data into an AI service unless your organisation has explicitly approved that processing.
Input: assignment brief and outcomes.
Expected output: a rubric with examples, moderation notes, and accessibility check.
Tool availability, model behavior, plan limits, and terms can change. Check each vendor's current documentation and your organisation's account settings before configuring a workflow.
Write one sentence describing the decision this work must support: a rubric with examples, moderation notes, and accessibility check. List what is confirmed, what is unknown, and who can approve a final version. Put source links beside material claims; generated text is never the source of a claim.
Create an input packet containing assignment brief and outcomes. Remove names, direct contact details, account credentials, and any material that is not needed for this task. For example, use a source link or internal record ID in place of a pasted confidential document. State what the output must not do: it must not make a decision about a person, publish automatically, or represent an uncertain statement as confirmed.
Handoff: give the bounded brief, input packet, and source folder to the person doing the first draft. That person should be able to explain why every included item is necessary.
For each criterion, write what beginning, developing, and proficient work looks like using observable evidence. Moderate the rubric against anonymised sample work with another assessor before learners receive it.
In Claude, ask for a structured first pass-not a final answer. A useful instruction is: “Use only the material below. Produce headings, an evidence table, open questions, and a draft. Label anything not supported by the material as needs verification.” Attach the source IDs next to the relevant paragraph or row.
Read the result line by line against the input packet. Delete claims that cannot be traced, and separate a useful hypothesis from a decision. The working output at this point should be a source-linked draft, not a polished asset.
Handoff: pass the draft, source links, and unresolved questions to the domain reviewer. Ask the reviewer to respond with corrections, missing context, and an explicit go/no-go on any sensitive claim.
Use Notion AI to organise the reviewer’s changes into three groups: required corrections, choices the accountable owner must make, and optional improvements. Do not ask the model to resolve a disagreement or infer approval. The accountable owner decides which version becomes the approved working copy.
Only then use Canva Magic Studio to make the requested output: a rubric with examples, moderation notes, and accessibility check. Keep the source-linked working copy alongside the formatted version. If a handoff changes format-for example, from notes to a deck, a prototype, or an automated message-compare the two versions for dropped qualifiers, altered numbers, and misleading visual emphasis.
If automation is part of the final step, configure a named human approval before anything is sent, published, or changes a record. Include a stop condition, an error alert, and a way to reverse a partial action without deleting the evidence trail.
Handoff: the accountable owner signs off on the final a rubric with examples, moderation notes, and accessibility check.
Before release, confirm:
AI can omit context, produce plausible errors, and reflect bias in the material it is given. This workflow is unsuitable for autonomous decisions about people, regulated advice, or irreversible actions. Start with a small sample and inspect usage-based charges, generation credits, storage, and paid-seat limits before scaling.