Feedback that is actually useful

Encouragement is not feedback

Ask a model what it thinks of your essay and you will get praise, then some gentle suggestions, then more praise. This is a consequence of how these systems are tuned, and it is worse than useless: it consumes your time and returns a false signal about quality.

The fix is to remove the conditions that produce it. Ask for something specific, adversarial and structured, and refuse the encouragement explicitly.

Act as a marker who has read two hundred of these and is behind schedule. Identify the three things costing the most marks, in order. Quote the sentence that demonstrates each. Do not list strengths. Do not rewrite anything. Do not be encouraging.

That produces feedback. The framing does the work: a specific reader, a specific job, a fixed number of items, evidence required, and the failure modes named and prohibited.

Ground it in the actual criteria

The single highest-value thing to paste alongside a draft is the assessment brief and the mark scheme. Nearly nobody does this, and it transforms the output.

Here is the brief, the mark scheme, and my draft. For each criterion, state which band this currently sits in and quote the evidence for that judgement. Then give the three changes that would move the most marks. Do not rewrite.

This works because it is grounded, and because it converts a vague quality question into a rubric-matching task, which is the kind of thing these systems do well on supplied text.

It is also a permitted use in nearly every institutional policy, since it produces no text for your submission. Check yours if unsure.

The order of feedback rounds

Getting feedback in the wrong order wastes it. Structure over sentences, always.

1. Structure. Topic sentences only. Is there an argument? Do the paragraphs depend on each other?
2. Evidence. Is each claim supported? Is any source doing more work than it can bear?
3. Objections. Is the counterargument the real one, at its strongest?
4. Sentences. Clarity, only after the first three are settled.
5. Formatting and references.

Polishing sentences in a paragraph you will later delete is the most common way students waste an evening. Do not edit prose until the structure is fixed.

Human feedback is still better, and there is more of it than you think

• Office hours. Chronically underused. Ten minutes with the person marking your work is worth more than any amount of AI critique, because they know what they are looking for and they will tell you.
• Ask a specific question. Not is this any good but is my third paragraph doing the work I think it is. Specific questions get specific answers and take less of everyone’s time.
• Writing centres. Most universities have one, free, staffed by people whose job is exactly this.
• A peer. Swap essays with someone on a different question. A reader who does not know the topic will find every place your reasoning was implicit, which is precisely what a marker will find.

Feedback on returned work

The feedback you already have and did not use: your marked assignments. Most students read the grade and close the file.

A better routine. Collect the comments from the last five pieces of work into one document, and look for what repeats. Almost everybody has two or three persistent faults (not enough analysis, unsupported claims, doesn’t answer the question) and fixing one of them raises every future mark.

A model is useful here: paste the comments and ask what patterns appear across them. That is analysis of supplied text, which is the safe category, and it surfaces things you have stopped seeing.

Feedback you give to others

Reading somebody else’s draft is one of the fastest ways to improve your own writing, because faults are far easier to see in another person’s work than in your own. You will notice the unsupported claim and the missing analysis instantly, and then find yourself noticing them in your own draft the following week.

Give it the way you would want it: name the two things costing the most, quote the sentence, and say what would fix each. Do not rewrite their prose.

The one to be wary of

Asking a model to predict your grade. It will give a number, it will sound confident, and it is unreliable: it does not know your institution’s standards, your cohort, or your marker. Students act on these numbers, decide a draft is fine, and submit.

Use it for which criteria are weakest and why, which is diagnostic and checkable against the rubric. Ignore the number.