AI Formative Assessment That Changes Tomorrow's Lesson
Dr. Steve Troyer · August 17, 2026
You give a 10-question formative assessment. 14 students miss number 4. Now what?
A score tells you how many missed it. It tells you nothing about why, and why is the only part that changes what you do in the morning. So don't ask AI to generate the questions. Ask it what each wrong answer should tell you, and you get back something you can plan from: the assessment, the answer key, a map from each wrong answer to the misconception behind it, and a reteach plan for tomorrow.
Start with what you might do next
Before you write a single question, name your options. There are usually 3: reteach it, give more practice, or extend.
The assessment exists to sort students into those 3 piles. That's its whole job. Any question that can't put a student in one of them is decoration, and decoration is why so many exit tickets give you a number and no next move.
What prompt do you use?
Prepare a 10-question formative assessment for [grade] [subject] on [objective]. Target these misconceptions: [list]. Include 2 items that require students to explain their reasoning. For each wrong-answer pattern, name the likely gap. Return an answer key and a short reteach guide for tomorrow.
Two parts of that are doing the work. Naming your misconceptions up front is what makes the wrong answer choices mean something instead of being 3 plausible wrong numbers. And asking for the likely gap behind each wrong-answer pattern is what turns a score into a decision.
What comes back?
- The student copy
- The answer key
- A map from each wrong answer to the misconception behind it
- Reteach, practice, and extension notes
The map is worth the whole exercise. A score tells you that 14 students missed number 4. The map tells you that all 14 picked C, that C is what you get when you set the ratio up upside down, and that this is one reteach and not a unit.
Anthropic shows standards-connected formative assessment with difficulty levels and an answer key as a Claude for Teachers workflow. The map and the reteach guide aren't a feature. They're something you ask for, which is why the prompt above is longer than “write me an assessment.”
Take it yourself before students do
All 10 questions, not a sample. It runs about 10 minutes and AI isn't trustworthy enough to not at least get your eyes on it.
- Is the answer key right?
- Is the reading load right for this class, or does the question test reading instead of the standard?
- Is anything ambiguous?
- Does it measure the standard, or does it measure test-taking skill?
My one warning
A bad question still gives you data; it's just confident and wrong.
If a wrong answer choice is ambiguous enough that a student who understands the concept picks it, that student now looks like they have the exact misconception the question was built to catch. The map will report it cleanly. The reteach guide will build tomorrow around it. And if you're keeping a class folder, that result goes in the folder and shapes your grouping for the next 2 weeks.
Nothing anywhere in that chain will flag it. It looks like evidence the whole way down, and it gets more authoritative at every step, because every step treats the one before it as sound.
That's why taking the assessment yourself isn't a diligence line here. It's the only place in the chain where the mistake can still be caught.
What is next
The 21-day Claude for Teachers series ended today, and the complete Field Guide is free. This keeps going weekly, and the next one is feedback: comments that sound like you instead of like a rubric. If you're new here, the class folder edition is the one this follows, and the setup edition is where to start.
What's the misconception you see every single year, no matter how you teach it? Reply and I may build the next example around it.
Every workflow in the series, with the prompts and the review steps, in the free Field Guide.
Get the Field Guide, freeCommon questions
- How do you use AI to write a formative assessment?
- Name the misconceptions you want to detect and the decision you are trying to make before you ask for any questions, then ask for an answer key, a map from each wrong-answer pattern to the likely gap, and a reteach guide. Asking only for questions and a key gives you a score with no next move attached.
- Can AI tell you why students got a question wrong?
- It can propose a likely gap behind each wrong-answer pattern when you tell it which misconceptions to target, which is what turns a score into a decision. The proposal is a hypothesis you check against student work, not a finding.
- What should you check before giving students an AI-generated assessment?
- Take all of it yourself rather than spot-checking: confirm the answer key is right, the reading load fits the class, nothing is ambiguous, and each question measures the standard instead of test-taking skill. It runs about 10 minutes for 10 questions.
- Does Claude for Teachers create formative assessments?
- Anthropic shows standards-connected formative assessment with multiple difficulty levels and an answer key as a Claude for Teachers workflow. The error map and the reteach guide are not a product feature; they are things the teacher asks for in the prompt.
- What is the risk of using AI-generated assessment questions?
- An ambiguous wrong-answer choice makes a student who understands the concept look like they hold the exact misconception the question was built to catch, and every step after that treats the result as sound. Working the questions yourself is the only place in the chain where that mistake can still be caught.