User guide
How to use this app
Five steps from a pile of answer scripts to one approved teaching decision. A full run on 50 responses takes about three minutes.
- 1
Enter the assessment
Type the exam question and define 2–5 rubric criteria — each with a name, what full marks looks like, and max marks. The rubric is what the AI classifies against, so sharper criteria give sharper diagnosis.
- 2
Add student responses
Paste responses (one per line, or separate long answers with a line containing ---), or upload a CSV with columns id,response. Minimum 5, tested up to 50. Or click Load demo data for a curated 50-response economics batch.
- 3
Analyze
One batched model call classifies every response by the belief behind it: Correct, Partially correct, Misconception, or Needs teacher review (anything under 60% confidence goes there automatically).
- 4
Read the diagnosis
The summary chips show the class split. Misconception clusters name each wrong belief with verbatim evidence quotes and a confidence score. The learning-gap map shows % mastery per rubric criterion, flagging the critical gap.
- 5
Decide
The AI proposes exactly one intervention with its rationale. You Approve, Modify, or Reject (with a reason). Every decision is logged. Draft marks exist in the detail table but are never final without you.
Input formats
Paste
PED is %ΔQd / %ΔP... Elasticity means demand falls when... --- A longer multi-line answer can sit between --- separators like this. ---
CSV
id,response R01,"PED is the percentage change in..." R02,"Elasticity is the slope of..."
The 3-minute demo path
| 0:00 | Load demo data → Analyze. 50 real-looking answers on price elasticity. |
| 0:30 | Summary chips: 18 correct, 21 partial, 11 misconception. |
| 1:00 | Open the top cluster: 5 students confuse elasticity with absolute change — their own words as evidence, with a confidence score. |
| 1:30 | Gap map: Interpretation is the critical gap at 43% mastery. |
| 2:00 | Read the AI's single proposed intervention and its rationale. |
| 2:30 | Click Approve: 32 of 50 students targeted instead of reteaching all 50. Decision logged. That decision is the product. |
Tip for presenters: add ?demo=1 to the URL to use the cached analysis and remove any dependency on live connectivity.
What this app will never do
- • Finalise a mark. Every mark on screen is a draft until a teacher confirms it.
- • Act on a student. Interventions run only after explicit teacher approval.
- • Use demographic data. Only the response text is ever analysed.
- • Hide uncertainty. Confidence is shown everywhere; low confidence goes to human review.
← Back to the app·How it's built, and the path to production →