Why assignment design plateaued
The early playbook was hopeful: make prompts personal, local, current, or weird, and the AI can't answer them. Teachers tried it at scale, and teacher forums spent 2023 through 2025 converging on the same disappointed conclusion: a student can paste any prompt, however personal, into a chatbot along with three sentences of personal context, and get a passable draft back. The model doesn't need to know your town; the student can tell it. Design changes what the student must type into the chatbot. It doesn't change whether they can.
Meanwhile the enforcement side collapsed on its own schedule: detectors proved unreliable, a major university cancelled Turnitin's AI detection, and a false-flag dispute became a federal lawsuit. So the design conversation matters more, not less. It just needs honest expectations.
Five tactics that still earn their keep
1. Anchor to class-interior material. Require engagement with the specific class discussion, a peer's argument, or your annotations on the shared text. A chatbot can fake familiarity with The Crucible; it can't fake what Maria said in Tuesday's seminar unless the student reconstructs it, which is itself most of the thinking.
2. Grade the middle, not just the end. Collect a thesis paragraph Monday, an outline Wednesday, a draft Friday. Each checkpoint is small, but faking a coherent trajectory across all of them takes more effort than writing honestly. The work product becomes a process, and a process is harder to outsource.
3. Keep some writing in the room. Even one in-class paragraph per essay gives you a sample of each student's real voice and pace, which makes anomalies in the take-home portions legible. You don't need a lockdown browser for this; you need fifteen minutes and a prompt they haven't seen.
4. Make revision the graded act. Hand back feedback and grade the delta: what changed, and did the change show judgment? AI is good at producing first drafts and notably weaker at defending specific revision choices in the student's own framework.
5. Talk to them about it. A two-minute conference ("walk me through how you built this paragraph") is the oldest integrity tool and still among the best. Its limit is arithmetic, not validity: two minutes times 150 students, repeatedly, is where it breaks.
The tactic the lists leave out: capture the process
Notice what tactics 2 through 5 have in common: they all try to make the writing process visible, indirectly, by sampling it. Checkpoints sample it weekly. In-class writing samples it spatially. Conferences sample it after the fact. Each one buys partial visibility at a real cost in your time.
The direct version is simpler: have students write where the process records itself. When the writing environment captures keystroke rhythm, pauses, revisions, and pastes as they happen, the question "did this student write this?" stops needing detective work. You look at the record of the writing happening, the same way the Wake County teacher who cleared a falsely accused freshman did with version history, only at full resolution instead of snapshots.
AI-resistant design tries to make cheating hard. Process capture makes honesty provable. The first is an arms race you maintain forever; the second is a record that just exists. You need less of the first once you have the second.
A five-paragraph essay assignment, redesigned
Take the classic: "Write a five-paragraph essay on a theme in To Kill a Mockingbird." The redesigned version costs you about ten minutes:
Prompt: "In our March 4 discussion, two students disagreed about whether Atticus is naive. Take a side in that specific disagreement, using one scene we did not discuss in class." (Tactic 1: class-interior anchor.) Structure: thesis paragraph due Tuesday in class, handwritten or typed in the room (tactic 3); full draft Friday; revision the following Wednesday, graded on the delta against your feedback (tactic 4). Medium: drafted in a process-capturing editor, so the draft arrives with its own writing record and you spot-conference only where the record raises a question (tactic 5, focused where it's needed instead of spread across everyone).
Nothing in that design is exotic. The difference is that when a submission looks wrong, you are no longer choosing between an unreliable detector score and an unwinnable he-said-she-said. The evidence already exists.
Drop the redesign tax
Manupropria is the process-capturing editor in that workflow: you create the assignment, students write at a shared link with nothing to install, and every submission arrives with a visual replay of how it was written, under a paste policy you set. The design tactics above still make your assignments better. The capture means you no longer need them to also be your evidence.