Teachomatic what to automate, and what not to

First-Draft Feedback: The Line

Generate a draft of the feedback, then edit it. This is the arrangement most teachers land on, it is sensible, and it is where most of the measured time saving comes from.

It also degrades quietly, in a way that has a specific shape and a specific tell.

For a workplace-oriented comparison outside education, further details offers another way to look at measurement, workload, behaviour, or accountability.

What the line actually is

Not a category of task. The line is the point at which you stop editing and start accepting.

Week one, you rewrite most of it. Week four, the drafts are reading well enough that you change a word and move on. Nothing was decided; the standard drifted, because editing is work and accepting is not, and the output is plausible enough that accepting feels reasonable every single time.

For an external perspective on teaching, assessment, and feedback, see TeachThought.

The tell is simple: if you cannot remember what you changed, you did not edit. That question, asked at the end of a marking session, is the whole quality control mechanism.

Why the drafts are so easy to accept

Because they are good at everything except the part that matters.

They are accurate about the work. They are correctly matched to the criteria. They are well-written, appropriately encouraging, and specific about the text in front of them. Nothing in them is wrong.

What they cannot be is specific about this student, because the tool does not have that information. It does not know you said this three weeks ago, that this is the first argument they have ever attempted, that they are coasting, that something is going on at home.

So the failure is not visible in the artefact. The comment reads fine. It only fails at the point of delivery, when a student reads it and correctly concludes that nobody wrote it for them — and you never see that moment.

The three-part test before sending

Fast, and it catches nearly everything.

Could this comment be attached to another student's work? If yes, it has not been edited enough.

Does it contain one fact about this student that is not in this piece of work? Their trajectory, a previous conversation, a habit. One is enough.

Would I recognise this as mine? Not vanity — if the voice is generic, the relationship is not in it, and feedback is a relationship over a term.

Where the draft is genuinely better than what you would write

Worth saying, because the point is not that human feedback is always superior.

At script eighty of ninety, a generated draft is better than what an exhausted human produces. Fatigue degrades feedback badly and unevenly, and the students at the end of the pile get the worst of it — usually the same students, if you always mark in the same order.

On mechanics, it is more thorough and more consistent than most of us.

On criteria coverage, it will not forget a strand you always forget.

For a subject outside your specialism, if you are covering, the draft may genuinely know more than you do about what a good answer looks like.

The honest position is that the draft raises the floor and lowers the ceiling. That is a good trade for the bottom of the pile and a bad one for the top.

The arrangement that works

Generate for everyone, edit hard for some. Full editing for the students where a comment will actually change something: the ones on a boundary, the ones who have shifted, the ones you are worried about. Lighter for routine work.

That is not cutting corners; it is where an experienced marker was already spending their attention. The difference is doing it deliberately rather than by whoever you reached first.

Add one sentence to every single one. Fifteen seconds, written by you, containing something only you know. This is non-negotiable and it is the entire defence against the drift described above.

Read the draft before the work occasionally. Reverses the anchoring and shows you what you would have said unprompted.

The wider version of the same drift

This pattern recurs across everything on this site, and it is worth recognising as a pattern.

Something plausible is produced. Checking it is work. The plausible version is nearly always fine. Standards slide by an amount too small to notice on any given occasion, and there is no moment where anything goes wrong.

The countermeasure is the same everywhere: a small mandatory human contribution that cannot be skipped without noticing. One sentence per student here. Fifteen scripts read properly when marking at scale. Five minutes on paper before generating a lesson plan.

Not because the human version is always better. Because a step you must actively perform is the only thing that makes the drift visible to you.

The short version