Why default feedback is useless
Models are tuned to be helpful and agreeable, which in a marking context means grading generously and phrasing criticism as suggestion. Without a standard to apply, "good" is undefined, so it defaults to something like "competent relative to average writing" — a bar your coursework has already cleared and which tells you nothing.
The second problem is vagueness. "Strengthen your analysis" is not actionable. What you need is "paragraph three states the evidence and never explains why it supports your claim; that's where the analysis marks were", and you only get that by asking for it in those terms.
The three inputs that change everything
| Input | Without it | With it |
|---|---|---|
| The actual mark scheme or rubric | Marks against a generic notion of quality | Marks against what your examiner rewards |
| An explicit strictness instruction | Generous, encouraging, unhelpful | Harsh, specific, occasionally uncomfortable |
| A demand for located criticism | "Could be deeper" | "Sentence 4 of paragraph 2 asserts without evidence" |
Most courses publish the rubric or the band descriptors somewhere — the module handbook, the assessment brief, the past paper's mark scheme. Finding it takes ten minutes and it's the difference between feedback and small talk.
The marking prompt
- 1
Give it the rubric first, verbatim
Paste the band descriptors or the mark scheme before your work. Then ask it to state which band your work falls in and quote the descriptor phrase that justifies it.
- 2
Assign a strict role explicitly
"You are a strict second-year marker. Award marks only where the criterion is clearly met. Do not give credit for implied understanding." That last sentence is doing a lot of work — implied credit is where generous marking hides.
- 3
Ask where each mark was lost, by location
"For every mark not awarded, quote the sentence or the absence responsible." Absence matters: often the loss is something that isn't there, and a model will only tell you if asked.
- 4
Ask what a top-band answer would have added
Not a rewrite — a list of the two or three moves missing. That list is your revision plan for the next attempt.
- 5
Then ask it to argue the other way
"Now make the strongest case that this deserves a higher band." Comparing the two passes shows you which criticisms are robust and which were the model being agreeable in a different direction.
- 6
Fix, then re-mark blind
Submit the revised version in a fresh conversation without mentioning the first. Carrying the history biases the second mark upwards.
What it marks well, and what it doesn't
| Marks reliably | Marks poorly |
|---|---|
| Structure — does each paragraph do a job? | Whether a subject-specific claim is actually correct |
| Whether you answered the question asked | Your specific department's unwritten preferences |
| Missing analysis after evidence | Originality — it can't know what's been said before |
| Unsupported assertions | Whether a source you cited says what you claim |
| Clarity, hedging, redundancy | Borderline band judgements, which are genuinely hard |
| Whether the conclusion matches the argument | Anything requiring knowledge of your cohort's standard |
The pattern: it's good at form and weak at content truth. Use it for the shape of the argument and verify every factual criticism against your own material — ungrounded models are confident about subject content in exactly the way that's hardest to catch.
For problem sets and calculations
Different job, different prompt. Don't ask whether the answer is right — ask where the reasoning diverges from the correct approach, and specifically not to give you the answer.
- "Here's my working. At which line does it first go wrong, and why?" This is the single most useful problem-marking prompt.
- "Don't give the correct answer" — otherwise you get the solution and lose the chance to fix it yourself, which is where the learning is.
- Verify arithmetic independently. Models make calculation errors, and a wrong correction is worse than no correction.
- Ask what the error suggests about a misconception. A slip and a misunderstanding need completely different responses, and the distinction is the useful output.
The self-deception risks
Three, worth naming because they're easy to fall into and hard to notice.
- Shopping for a better mark. Re-asking with a slightly different prompt until it agrees with you. If you're on the third attempt, you have your answer already.
- Accepting the rewrite. Asking it to fix the paragraph means you've learned nothing and, depending on your institution's rules, may have crossed a line — see is using AI for homework cheating.
- Treating it as the mark. It's a rehearsal, not a prediction. Your actual marker has a cohort, a house style and opinions the model can't know.
Where the real value is
Not in the mark. It's in the loop speed: you can write, mark, revise and re-mark three times in an evening, where waiting for a tutor's feedback takes two weeks and arrives after the next assignment is due. Three iterations on one essay teaches more about what a rubric rewards than three essays marked once each.
Which means the best use is on work nobody will mark — practice essays, past-paper answers, the second attempt at something already returned. That's the material that would otherwise generate no feedback at all, and it's where the difference between a good and an average technique gets built. Real tutor feedback remains more valuable per instance; you just can't get much of it, and office hours is how to get the most from what's available.