Building a rubric and calibrating grading with your TAs
Faculty rate grading as the task AI does worst, yet nearly half of grading-related AI conversations hand the judgment over anyway — the exact pattern researchers called concerning. The defensible version: use AI before grading starts, to build the rubric, the comment bank, and a calibration packet that gets you and your TAs scoring consistently.
You are an assessment designer for university courses. Create a grading kit for this assignment. Assignment: {{assignment_description}} Course and level: {{course_level}} Skills to assess: {{focus_skills}} Output three parts: 1. A four-level analytic rubric (Excellent / Proficient / Developing / Beginning) with one row per skill. Descriptors must describe observable features of the work in 25 words or fewer per cell, with no overlapping language between adjacent levels and no vague qualifiers like "adequate." 2. A feedback comment bank: for each skill, two sentence starters for work that meets the standard and two for work below it. Every below-standard comment must name one concrete action the student can take in revision. 3. A calibration packet for my TAs: write three short synthetic response excerpts — invented by you, not real student work — at clearly different quality levels, each with a suggested rubric score and a one-line rationale, so we can norm our grading before we start. Constraints: assess only the skills I listed — do not add criteria. Keep all rubric language student-readable, because the rubric will be published with the assignment. Do not ask me for real student submissions.
Fill in your details and the prompt updates live — then copy.
**Use of evidence — Proficient:** Cites four or more course readings accurately; each source is connected to the recommendation with at least one sentence of analysis. **Use of evidence — Developing:** Cites readings, but sources are summarized rather than used; connections to the recommendation are asserted, not shown. **Comment bank (below standard):** "Your Chen citation is accurate — now add two sentences showing how her findings support congestion pricing specifically." / "Pick your strongest source and move it into the recommendation section, where the evidence is currently thinnest." **Calibration excerpt B (suggested score: Developing):** "Air pollution is a serious problem that many scholars have studied..."
The full workflow
- Build the rubric before the assignment goes out, and publish the student-facing version with it.
- Delete any rubric row you wouldn't genuinely score differently across four levels.
- Run the calibration meeting with your TAs using the synthetic excerpts, and adjust descriptors where scores diverge.
- While grading, pull comment-bank language freely — but make every scoring decision yourself.
Watch out for
Faculty in Anthropic's survey rated grading as AI's least effective use, and researchers flagged that 48.9% of grading conversations delegated the judgment to the model. Use AI for the language of feedback, not the grade.
FERPA: if you paste student submissions in for feedback drafting, strip names, IDs, and headers first — and prefer an institution-licensed tool covered by a data agreement over a personal consumer account.
Where this comes from
Every use case on this site is grounded in real reports from working professors — not invented by us.