Proposed solution: Agentic rewriter

Keep a history of the teacher's actions. Send it to the LLM with the grading task and feedback on each rewrite. The model adapts the wording to the teacher's style without changing the meaning or scores.

Teacher-written text matters more than AI-generated text. Newer examples and instructions take priority when preferences change.

What goes into the history

Illustrative UI examples.

Feedback written by the teacher

When the teacher saves their own feedback, record it with the assignment, rubric, and student submission.

Grading

Feedback

Correct setup. Check the last division: 18 ÷ 3 = 6, not 5.

Save feedback

AI feedback edited by the teacher

Record the original AI feedback and the teacher's saved version, not every keystroke.

Grading

AI text

~~Your setup is correct, but the final calculation contains an error.~~

Your edit

Correct setup. Check the last division: x = 6.

Save feedback

Writing instructions changed Optional

Add a freeform "Writing style" setting. Record each saved change in the history.

Writing style

How you write

Keep feedback short. Avoid generic praise. Point to the specific step that needs correction.

Save instructionsCancel

Rewrite requested by the teacher Optional

Let the teacher ask for a change in plain language. Record the request, the suggested edit, and whether they accepted, rejected, or modified it.

Rewrite feedback

What should change?

Refer to the student in third person.

Rewrite using AI

Suggested edit

Current

~~You~~ expanded the brackets correctly, but ~~your~~ final division is wrong.

Suggested

The student expanded the brackets correctly, but their final division is wrong.

AcceptEditDismiss

Context compaction

When the history reaches a limit, for example 100 messages, summarize the older events and keep the latest 50 unchanged. Keep the system prompt fixed.

Costs

Roughly $0.04 per edit for history input.

Based on the earlier estimate of 2,000 tokens per event, cached history, and compaction from 100 to 50 messages. Current task input, output tokens, and summarization cost extra.

Implementation

Frontend

Add the optional writing instructions and rewrite request controls.
Send saved feedback, edits, instruction changes, and edit decisions to the backend.

Backend

Store a history for each teacher and append events received from the frontend.
Send the history, grading context, and feedback to ML when a rewrite is requested.

ML

Add an endpoint that rewrites feedback using the teacher's history.
Summarize older events when the context reaches its limit.

Libraries

No additional libraries are required. Reuse the existing storage and LLM connector. No RAG, vector database, or agent framework is needed.