ChatGPT for Teachers in 2027: Build a Grading Assistant That Uses Your Rubric and Waits for Your OK

Do the math with me. Four sections of ninth-grade English, 30 students each, one argument essay every three weeks. That is 120 essays per cycle. At 12 minutes each (read, mark, score four rubric rows, write the comment the student will actually read), you are looking at 24 hours. Three full workdays, outside class, every cycle. That is why the essay every three weeks quietly turns into an essay every six, and students reach the state writing test having written five argument pieces all year.

This post shows how to set up ChatGPT for teachers as a grading assistant: it reads each essay against your rubric, quotes the lines behind each score, flags where it is unsure and hands you a draft. You review, change what you disagree with and release it. In our estimate, time per essay drops to 5 or 6 minutes. The goal is not to take you out of grading. It is to take you out of the mechanical part of it. If you want the bigger picture first, the guide on AI for educators covers the rest of the teaching week.

What a grading agent is (and what it is not)

"Agent" has become a marketing word. Here it means something simple: a workflow that receives an input (the essay), runs a few steps on its own (reads it, applies the rubric, writes the comment, logs it in a sheet) and stops at a human checkpoint before anything reaches the student. The difference from "open ChatGPT and paste the essay" is that you do not paste anything. Students submit, the flow runs, and you show up to approve.

What it does not do, and should not do: give a final grade, send a comment to a student without you reading it, or grade work in a subject where you have not written a rubric.

The rubric the agent needs

The agent is only as good as the rubric you give it. For this post I used a four-row argument rubric aligned to the Common Core writing standard for grades 9-10 (W.9-10.1), scored 1 to 4 on each row. If your department or state test uses a different rubric, use that one. The structure of the setup does not change.

That is why the system prompt below is long and a bit boring. It forces the model to be suspicious about evidence and to check the counterclaim before scoring reasoning.

Two ways to build it: no-code and automated

Option 1: a custom GPT or a Gem (no code, 30 minutes)

In ChatGPT you create a custom GPT with the instructions below and attach your rubric as a PDF. In Gemini the equivalent is a Gem. You still paste each essay, but you never repeat the instructions or the rubric, and the output always comes back in the same format. This is enough for up to about 80 essays a month. If your district gives you ChatGPT or Gemini through a school account, build it there.

Option 2: an automated flow with Make or n8n (one afternoon)

Here students do not even see the AI. The path looks like this:

1. Input: students submit through a Google Form (typed text or a photo of handwritten work). Identification field: section and student ID number, no name.
2. Reading the photo: if it came as a photo, a model with vision transcribes it and marks unreadable words with [?]. Teenage handwriting is the weak point: check the transcription whenever there are more than three [?].
3. Grading: the text goes to the model with the system prompt and the rubric. The output comes back in a fixed format (JSON) with a score per row, quoted lines and a confidence level.
4. Document: the flow creates one Google Doc per essay with the text, the comments and the provisional score, in a "To review" folder.
5. Sheet: it logs section, ID, provisional scores and a "low confidence on" column, so you know where to look first.
6. You review: open the Doc, change what you disagree with and set the status column to "approved".
7. Return: only approved feedback goes back, through Google Classroom or your school email.

I would build it in Make if you are doing this alone: the interface is friendlier and the free plan handles a test with one section. n8n makes sense if your school has someone in IT who can host it. Either way, this version touches student work automatically, so it has to run on accounts your district approves (more on that below).

The agent's instructions (system prompt)

This is the core of the agent. Copy it into the custom GPT or the AI step of your flow and adjust the last part to the way you write comments.

You are an experienced 9th-grade English teacher scoring argument essays. Score the essay below with the attached 4-row rubric (Claim, Evidence, Reasoning and organization, Conventions), 1 to 4 per row.

Required rules:
1. For each row, quote the lines from the essay that justify the score.
2. Claim: check first whether the essay takes a position on the exact prompt "[EXACT PROMPT]". If it only discusses the topic, say so clearly.
3. Evidence: for each piece of evidence, say whether it comes from the provided sources [SOURCE LIST] and whether it is connected to the claim. If you cannot match a quote or statistic to a provided source, mark it "unverified" instead of accepting it.
4. Reasoning: say whether each body paragraph adds a new reason, and whether there is a counterclaim with a rebuttal. Only then score.
5. Do not round up. If you are between two levels, choose the lower one and explain.
6. Give a confidence level (high, medium, low) for each row.
7. Write two blocks: "For the teacher" (technical, short) and "For the student" (4 sentences max, in the voice of a teacher who is rooting for them: one thing that works and the two changes that would raise the score most).
Use American English. Do not use the student's name.

Real test: a ninth-grade argument essay

Prompt used: "Should schools ban cell phones during the school day?" with two short provided sources (a news article on a district phone ban and a survey of teachers). Below are the introduction and conclusion of a five-paragraph essay written for this test in a typical ninth-grade style (the body paragraphs are left out to keep this short):

"Cell phones are a big part of teenagers lives today. Some people think they should be banned in school and some people think they shouldn't. In this essay I will talk about why phones are distracting and why students need them for emergencies. [...] So overall, phones are good and bad, schools should make rules so students don't get distracted, that way everyone is happy."

Output generated with the system prompt above (full essay as input):

Claim: 2 (high confidence). The introduction announces both sides ("some people think they should be banned... some people think they shouldn't") without taking a position. The conclusion leans toward "make rules" but does not say whether phones should be banned during the school day.

Evidence: 2 (medium confidence). Paragraph 2 quotes the teacher survey correctly ("most teachers reported that phones were a major distraction") and cites it. Paragraph 3 says "studies show phones lower grades by 20%": this statistic does not appear in either provided source. Marked unverified.

Reasoning and organization: 2 (medium confidence). Paragraph 2 argues phones distract; paragraph 3 argues the same point with a new example. Paragraph 4 raises emergencies but never answers it. No rebuttal.

Conventions: 3 (high confidence). Missing apostrophe in "teenagers lives". Comma splice in the last sentence ("...rules so students don't get distracted, that way everyone is happy"). Otherwise controlled sentences.

Provisional total: 9 out of 16.

For the teacher: check the Evidence row. If the 20% figure came from a source you discussed in class, it could move to 3.

For the student: Your paragraph using the teacher survey is your strongest; you picked a good quote and explained it. The change that would raise your score most is in your first paragraph: pick a side and say it in one sentence. Then use paragraph 4 to answer the emergency argument instead of just mentioning it.

What I liked: it caught the unverified statistic, which is exactly what a tired teacher misses on essay 87. Where I disagreed: I kept Claim, Evidence and Reasoning as scored, but dropped Conventions to 2 because of a second comma splice in paragraph 3 that the model missed. That is the job of the review step. It took me about 4 minutes to agree, fix one score and release.

How much it costs per month

Put that next to the time. If the review step saves 6 minutes per essay on 240 essays, that is 24 hours a month. Even at a modest hourly rate, the tool pays for itself in the first afternoon.

A second agent: next week's lesson plans

Once the grading flow works, the same structure handles planning. You fill out a form with four fields (standard, number of days, what the class struggled with this week, any events on the calendar) and the flow drafts the week in your template in a Google Doc. It does not decide what to teach. It saves the hour of formatting you spend every Sunday. The prompts I would use for that are in the post with AI prompts for teachers.

When it is NOT worth building

• Fewer than 40 essays a month. A saved prompt in ChatGPT is enough. The automated flow is overkill.
• Your district has not approved any AI tool for student work. Then Option 2 is off the table, and Option 1 only works with anonymized text you paste yourself. Ask your tech department before building anything.
• Elementary grades. Short answers and handwriting from young students make transcription unreliable, and COPPA adds rules for children under 13.
• Your rubric is in your head. If you cannot write it down in four rows with examples, the agent will invent one. Write the rubric first.
• Creative writing. Poems and personal narratives depend on voice. The AI tends to reward safe writing.

Common mistakes when building your first agent

• Letting it send feedback straight to students. The approval step is the whole point. Without it, one wrong score in front of a parent costs you more than the time saved.
• Skipping the source list. If the model does not know which sources were provided, it accepts any statistic that sounds plausible.
• Using names in the form. Student ID and section are enough, and they keep names out of the AI step.
• Never calibrating. Grade the first 10 essays yourself, compare with the agent and adjust the prompt where you disagree. Repeat once a semester.
• Not telling students. One line in your syllabus ("I use an AI tool for a first read; every grade and comment is reviewed by me") avoids an awkward conversation later.

FERPA and the right account

Student essays tied to an ID are education records under FERPA. A school can share them with an outside service only under specific conditions, which in practice means the district has a data agreement with the vendor. So the rule is simple: the automated flow runs only on accounts and tools your district has approved. Your personal ChatGPT account is fine for planning and for anonymized samples you paste yourself, and that is it.

Three more habits: turn off training on your data in the tool settings, strip names and any other identifiers before text reaches the AI, and keep the Docs in your school Drive, not a personal one. If you teach students under 13, also check COPPA and your district's rules before any student interacts with a tool directly.

If you would rather not build any of this

Building the flow takes an afternoon and a bit of patience with Make. If you would rather have the grading assistant and the weekly planner ready to use, already set up with the approval step and the rubric format, there is a ready-made version below. And if you want to try the prompts by hand first, the AI for educators guide is the best place to start.

Advertisement

Want to skip this part?

Ready-made AI agents for teachers: nothing to build, no prompts to write, they already work for you.

• A first read of your students' essays with your rubric, waiting for your approval
• Next week's lesson plans drafted from a four-field form
• Accommodated quizzes and activities with the answer key solved

I want the ready-made agents

Step by step

1. Write your rubric in four rows with examples of each level
2. Create a custom GPT or Gem with the system prompt and attach the rubric
3. Grade 10 essays yourself and compare with the agent's drafts
4. Adjust the prompt where you disagree
5. If your district approves it, automate with Google Forms, Make, Docs and Sheets
6. Release feedback to students only after you mark it approved