AI prompts for feedback and grading comments

Written for teachers. Rubric-anchored comments in your voice, specific to the work — it drafts the feedback; the grade stays yours.

• 01 Ten sections, from the rubric you translate before the stack to the file you still open in March
• 02 30 fill-in prompts — comments on one piece of work, the hard truth said well, the conference, the paper that missed everything
• 03 30 follow-up steps that take a first draft to something you would hand a fourteen-year-old
• 04 30 failure modes, each with the prompt that repairs it
• 05 An 83-page manual, a printable quick-reference page, and a plain-text file every prompt copies cleanly from

The first prompt in the kit does not ask a model what it thinks of a student’s essay. It asks you. You paste the rubric criteria and one piece of work with the name block stripped, then state your call in a phrase — strong thesis, evidence thin in body two, works cited sloppy — and what comes back is that call turned into comments the student can act on. Each one quotes an actual line from the paper, ties it to a criterion, and hands over one concrete next move. No score appears anywhere, because no prompt in the kit asks for one.

In the run printed in that section, the teacher’s judgment was four words: one example repeated twice. The comment that came back set the student’s two sentences side by side so he could see the repeat, named the criterion it fell under, and pointed at a line elsewhere in his own draft as the place to go find a second kind of evidence. That is the shape the whole kit aims at — feedback that could only belong to this paper. A comment that would survive being stapled to somebody else’s essay is not feedback, and students learn to skim past it.

The follow-up chain is the rest. One step rewrites every comment so a student at that grade reads it in one pass. One adds an opening line that names something real, quoted from the work — and says plainly when there is nothing honest to praise rather than manufacturing it. One turns the comments into talking points for five minutes at the student’s desk. The ten sections work the same way: the rubric translated into stems, comments on one piece of work, a voice card built from your own past comments, the reusable entries you write after the first five papers, the hard truth said well, the volume workflow for a stack of thirty, the writing conference, peer-review sheets students fill in, the paper that missed everything, and the file you keep.

Feedback has its own failure modes, and they are not the ones people warn you about. The kit names thirty of them before you meet them, and gives you the wording that repairs each.

The one that breaks the promise is the model grading anyway. You told it the verdict was yours, and the comments arrive with a bonus “Overall: B+” or “4/6” on the end, because a model trained to be helpful completes the pattern. The grade a model invents is the one a student quotes back at you in November. The expensive one is quieter. Somewhere around paper twenty your grading notes go terse — “fine,” “ok,” a dash — and those papers come back with full comments anyway, complete with a quoted sentence in quotation marks that the student never wrote. It reads exactly like the twenty-nine real ones. Its neighbour is thinning: the first ten comments are specific and by paper twenty-four they are “good work, add more detail.” The bottom of the stack is not a worse group of students. It is just later.

The rest are familiar once named. Comments written about the student instead of the work — careless, lazy, struggles with — which nobody can revise before Friday. A praise sandwich rebuilding itself around the gap so the gap ends up in the soft middle. Six problems found where you named one. A model that quietly rewrites the student’s paragraph and hands back the finished version, which turns the redo into a copying exercise. And the model volunteering, unasked, that a polished paper “may be AI-generated.” That one gets its own repair, because it is not the model’s call and it never will be.

The free starter card carries the four rules and ten cross-cutting prompts at principle level — enough to get a usable first draft out of any model. It does not carry Kit 05’s judgment brackets, its follow-up chains, its thirty failure modes and repairs, or the quick-reference page. That line is deliberate and it does not move.

The compliance rails here are structural, not stapled on. Nothing in the kit grades: every drafting prompt takes your assessment as an input, so a prompt with that slot left empty has nothing to say — which is exactly what it does in testing, refusing to draft and asking for your real call. Nothing accepts “is this AI” as a job. Where a prompt does take student work, it takes one piece per chat with the name block and any classmate named in the text removed, and the instruction sits inside the bracket where you cannot miss it. Whole sections never touch student work at all — the rubric section uses your rubric, the peer-review section uses your criteria, and the stack-of-thirty workflow runs on a page of your own notes, numbers not names. The papers stay on your desk.

Every prompt in the kit was run in the exact wording that ships, against Claude, in August 2026, and the sample outputs printed in each section are excerpts from those runs rather than illustrations written for the page. Nothing outside the Claude family has been tested, and this kit does not claim it. The repairs work by adding constraints rather than by exploiting any one model’s quirks, which is the honest version of portability.

Teacher Kit 05 · $7 · 83-page manual · on sale now