Canvas Grading Assistant

Your key is encrypted at rest on this device and sent only to Anthropic when you grade — never to any other server. If this machine is shared, rotate the key when you're done. Treat a leaked key like a leaked credit card — revoke it on console.anthropic.com.

Off by default. When on, the AI marks passages worth a second look. These flags are observations for your judgment, not evidence — writing-style signals have many benign causes. Nothing is shown to students, and you decide what, if anything, a flag means.

When a file has too little readable text (a scan, a photo of a worksheet, handwriting), the grading assistant sends its pages as images instead. Typed PDFs and Word files are still read as text. Page images use more API credits than text. Reading handwriting is less reliable than reading typed text, so these grades always wait for your review and are never applied automatically.

Limit the number of pages (or photos) sent as images, to control API costs. If a submission is longer, you are told which pages were left out.

Fill the grade input with AI-suggested score after grading.

Automatically post feedback to submission comments.

Every grade names the model that produced it. Changing the model changes grades; re-grade a few students you already know before switching a whole class.

How much the model thinks before it scores. Reasoning is always on for Claude 5 models; on older models it can be turned off below, which makes grading cheaper but less careful.

Off: the model answers directly at temperature 0, so the same work gets the same grade every time. Claude 5 models ignore this and always reason.

These rules go to the model with every grade and do not change per teacher. Your strictness setting adds the scoring policy shown at the end. The full prompt, with your preferences and the assignment's rubric, is under "Grading prompt" below.

Hard ceiling on Anthropic spend per day, enforced before every grade (single and batch). Resets at midnight. Blank = unlimited.

"Downgrade" keeps grading past the cap using a cheaper model. Downgraded grades are flagged and held for review (never auto-applied).

Model ID to use once the daily cap is reached, e.g. claude-haiku-4-5.

Absolute daily ceiling — even downgraded grades stop here. Blank = stop at the cap above.

Clears the session tallies; today's cost-cap meter keeps counting and grading history is untouched.

Verbose console output across the extension, including prompt cache hit/miss rates. Leave off unless you're troubleshooting.

Downloads a JSON file with your recent grading events, session totals, and settings (API key redacted). Helpful when filing a bug.

This is the prompt sent to Claude. Edit it to customize your feedback style.

Using default prompt based on settings above.