2026-08-25 · Guide

Bots for Educators: Planning and Feedback Support

The teaching week is contact hours plus everything around them

A secondary teacher with a full timetable stands in front of classes for something like twenty hours a week. Those hours are fixed and they are not the problem. The problem is the ring of work around them.

Planning five lessons across two year groups, most of which is not invention but retrieval: finding the diagram you used last year, rebuilding the starter activity you cannot locate, making the version of the worksheet for the three students who need it. Marking, which arrives in lumps and eats a weekend. Parent and admin email. The trip forms. Cover work written at 22:40 for a colleague you are covering for tomorrow. The department meeting. The data drop. The spec change nobody circulated.

Look at that list honestly and it splits in two. One half is retrieval, formatting, scheduling, and watching for changes, all of which have outputs you can verify in a minute. The other half is judgment about specific young people, and it is regulated, consequential, and the actual job.

Almost everything that goes wrong when a teacher adopts bots comes from mixing the two halves. So this guide keeps them apart from the first paragraph: the roster below sits entirely outside the gradebook, and the section on feedback explains exactly where the line runs and why.

Six bots that sit outside the gradebook

Not one of these needs a student's name, a piece of student work, or a mark. That is a design choice, not an accident.

That last boundary is narrower than the catalog default on purpose. A general inbox bot reads the whole mailbox. A teacher's mailbox contains safeguarding threads, pastoral notes, and messages about individual children, so the scope becomes logistics mail only and the exclusion goes in the charter as a rule rather than a preference.

Sort every classroom task by what a wrong answer costs a student

That roster is six jobs. You will think of a seventh by Friday, and the useful question is not whether a bot can do it but what happens when it does it wrong. Sort by that and most cases answer themselves.

Read the third column before the first. Every task where a bot is straightforwardly useful is one where nobody will ever ask you to explain the output. The moment the answer to "who justifies this" is a person, in a meeting, about a named child, you have left the range where a bot helps.

Student data is not yours to paste

State this once and get back to work. Student information is legally protected in most places. In the United States that is FERPA. In the EU and the UK it is the GDPR, where a school is typically the data controller and any processor it uses has to be under a contract it approved. Other jurisdictions have their own equivalents.

None of this is legal advice, and the practical instruction is simpler than the law: check your institution's policy before you paste anything, because almost every school, college, and district already has one. There is usually an approved-tools list, a data protection officer, and a process, and the process exists precisely so that individual teachers do not have to make this call alone at 22:40.

The reason a bot runtime lands squarely inside that policy is worth understanding rather than taking on faith. A bot does its work on a cloud computer, which is a third-party system by any definition your institution uses. In Grok Bot's case the documentation is specific about what that computer is: it is assigned to your user account and not to an individual bot, every bot you create gets its own screen on that same machine, and browser cookies, signed-in sessions, files, and command-line credentials are shared across all of them. The docs put it flatly: do not use separate bots as a security boundary.

Two consequences follow directly. First, the tidy-sounding plan of making one bot for student work and keeping it apart from everything else does not describe anything real. Second, deleting that bot does not remove the files or the browser sessions it left behind on the shared computer. Work uploaded there is work that stays there until you remove it deliberately. As of writing there is also no audit view of bot actions, which means you cannot later demonstrate what was accessed, and a data question you cannot answer is its own problem. The shared computer security guide covers exactly what is shared and how to clean up.

Sort your inputs before you paste, not after

The section above is about permission. This one is about what you are holding, because most of us underestimate how much of the material on a teacher's desk carries a person inside it.

The right-hand column is the actual skill, learnable in an afternoon. Nearly every classroom task can be described without the person in it. "Year 10, mixed attainment, third lesson on moles, most of them stopping at the conversion step" carries everything a planner needs and identifies nobody.

Where a task genuinely cannot be described without a named student in it, treat that as the signal that the task is yours rather than that you need a better prompt.

Curriculum and spec changes are a watcher's job

A cheerful one, and the highest-value thing on the roster that has nothing to do with students.

Awarding bodies, ministries, and districts update specifications, sample assessment materials, subject guidance, and grade boundaries on public web pages, usually without telling anyone in a way that reaches a classroom teacher. Departments find out in September that something moved in April.

That is a watcher's exact shape: a public page, a schedule, and a diff. The useful report is not "the page changed", because a cookie banner or a footer year will trigger that. It is the changed clause, quoted, with the date it appeared and a one-line note on which of your units it touches. Point it at your subject's spec page, the sample assessment page, and any guidance page you have ever had to reread, run it weekly, and read it in two minutes on a Friday.

Nothing about it needs a student, a login, or a write permission. It reads public pages and never interacts with them, which is also the boundary the catalog listing already declares.

Draft the feedback, and never decide the outcome

Here is the line, stated once and plainly.

A bot must not grade, rank, or determine an outcome for a student on its own. Not a mark, not a grade boundary decision, not a set placement, not a predicted grade, not a pass or fail. Drafting feedback that a teacher reads and edits before it reaches the student is the defensible scope. Everything past that point is an outcome, and outcomes belong to a person who can be asked to justify them.

That framing is not squeamishness. It is the same standard every school already applies to a teaching assistant, a trainee, or a supply teacher: they can prepare, they can suggest, and the assessor of record signs. A bot gets the same status as any other unqualified helper, which is a perfectly useful status.

If you do work within that scope, the mechanics matter. Any drafting against real student work is a data question first, so it happens only if your institution's policy allows it and in the way that policy specifies, which for many schools means an approved platform rather than a general-purpose bot runtime. Where it is allowed, the bot works against a rubric you wrote, it never sees or invents a mark, and it never sees the student's name or previous grades. Feedback drafted without knowing whose work it is stays feedback about the work.

Treat the edit as the reading, because that is what it is

The obvious reading of "the teacher edits it" is that the edit is a safety check, a rubber stamp on a draft that is basically fine. That reading is wrong, and it is worth being precise about why, because a stamp is exactly what the arrangement decays into if nobody names the failure.

The edit is where you read the work. That is the whole thing. Marking is not primarily the production of comments, it is the occasion on which a teacher finds out what a class actually understood. If a draft comment lets you skip that, you have not automated the writing, you have automated the reading, and the reading was the assessment. The comments were always a byproduct.

The edit is also where the specificity comes from. A model working from a rubric produces the median observation: the conclusion needs more evidence, the method could be clearer. Those sentences are true about almost any piece of work, which is why they change nothing. What changes behaviour is the sentence only you can write, because you know this is the third time this student has stopped at the same point, or that this paragraph is a genuine leap for them, or that the class misconception you taught against on Tuesday is sitting right there in paragraph two.

And the edit is where you can defend it. You may have to explain a piece of feedback to the student, to a parent, to a moderator, or to a head of department. A comment you have not actually made is a comment nobody made. That is uncomfortable in a meeting and it is worse for the student, who is entitled to the judgment of a qualified person.

There is a fairness dimension too. A bot applies a rubric with consistent tone and inconsistent substance, and its errors do not distribute evenly. It can be confidently wrong about an unusual answer, an unconventional structure, or a student writing in an additional language, and those errors are invisible in aggregate while being significant for the individual. The teacher edit is the only place that gets caught.

The practical test: if reading the work and rewriting the draft takes you as long as writing the comment from scratch, use the bot for planning instead. It has plenty to do there.

The strongest case for AI marking, and what it actually wins

The objection to all of that is not technical, and it deserves a straight answer rather than a restatement of the rule.

It goes like this. Marking is the largest single driver of teacher workload and one of the main reasons people leave. Saying that the job eating your weekends is the one job which must stay manual is a comfortable argument from someone without 120 books to get through by Monday. If a bot produces a usable comment in four seconds, refusing on principle chooses the teacher's exhaustion over the student's feedback, and a student gets better feedback from a tired teacher with a tool than from one writing nine words at 23:00.

The part which is right is the part about volume, so here is what it wins.

It wins on whole-class patterns. Twenty responses to one question, described without names attached, produce a decent account of where a class went wrong. That is analysis of a set rather than assessment of a person, and it is the most useful thing here, because it changes what you teach on Tuesday instead of what you write in a book.

It wins on your own comment bank. Marking is slow partly because you retype the same six paragraphs in slightly different words all term. A bot that helps you write those six once, properly, and then helps you choose between them, has automated the typing and left the judgment alone.

It wins on low-stakes formative work, where a wrong comment costs a student one wasted attempt and a correction on Thursday.

It stops winning where the reading disappears, and volume does not answer that. With 120 books and no time, the honest response is fewer and better-designed assessment points rather than the same number processed faster, because faster processing tells you nothing about the class. The exhaustion argument also cuts the other way: a rubber stamp is the fastest available way to spend a Sunday and learn nothing.

Start with the planner, the one bot that never sees a name

This is the highest-value bot on the roster and it never sees a child's name.

ROLE
You are my lesson planning assistant. You prepare materials I teach from.
You are not involved in assessment of any kind.

TRIGGER
Every Thursday at 16:00, prepare next week's plans from my scheme of work
file. Also run on demand when I name a topic.

INPUTS
- My scheme of work file (topics, sequence, dates, class codes only)
- My existing resource folder, to reuse before inventing
- Public subject resources, exam board specs, and sample assessments
- The misconception list I maintain per topic

OUTPUT, per lesson
1. OBJECTIVE: one sentence, in student-facing language.
2. PRIOR KNOWLEDGE ASSUMED: three bullets, and one check question for each.
3. STARTER: 5 minutes, low floor, no printing required.
4. SEQUENCE: 4 to 6 steps with rough timings. Mark where I talk and where
   they work.
5. MISCONCEPTION PROBE: one question whose wrong answers are diagnostic,
   with what each wrong answer tells me.
6. EXIT TICKET: two questions, one recall, one application.
7. REUSE NOTE: which of my existing resources this replaces or needs, by
   filename. Say NONE FOUND rather than describing a file that is missing.
8. DIFFERENTIATION: a simpler entry point and an extension, described by
   task, never by naming or describing a student.

RULES
- Never write an assessment mark, grade, band, or predicted outcome.
- Never rank, sort, or compare students in any way.
- If my scheme of work is ambiguous about sequence, ask. Do not guess.
- Cite the spec point each objective maps to, or say UNMAPPED.

BOUNDARY, absolute
You never read, store, receive, or process any student name, student work,
mark, grade, attendance record, behaviour record, pastoral note, or
safeguarding information. If any input contains one, stop, tell me which
input, and process nothing from it. This holds even when I ask.

The last sentence is the one that does the work. A boundary that you can lift in a hurry on a Sunday night is not a boundary, and Sunday night is exactly when the temptation arrives. If you want the phrasing that makes a rule like that hold, writing a boundary that actually constrains a bot covers how to word it so it cannot be argued around.

Keep these six jobs away from every bot you run

The opinionated list. These are lines that should still hold in a year, and several of them are lines your employer has probably already drawn for you.

Grades, marks, bands, and predicted grades. Any number or letter that lands on a record. This includes the version where the bot suggests and you agree without recalculating, because that is the same act with an extra click.

Anything that determines a student's path: set placement, intervention selection, entry tier, and reference or recommendation letters. A reference is a professional judgment with your name on it and a real effect on a young person's next year.

Safeguarding and pastoral concerns. These go to a designated person, by the named route, immediately, with no intermediate system involved. There is no version of this that is a triage problem.

Messages to parents about a specific child. Draft the trip logistics email all you like. The message about how a child is doing is a relationship, and often a difficult one, and it is read very carefully at the other end.

Detecting cheating or AI use. Accusation tools are unreliable in a way that falls hardest on students who write unusually, and an accusation is an outcome.

Anything that requires the student to believe a person read their work. If they suspect a machine is reading it, the reason to write for a reader disappears, and so does most of the point of the exercise.

Test the planner in the week you are ill, not in October

The test for a planning bot is not whether the plans are good in October, when you are fresh and checking everything. It is whether they hold up in the week you are ill, covering two colleagues, and running a trip.

So check it in two cheap ways. Once a fortnight, take one lesson the bot planned and one you planned, and compare how much you changed on the morning. If the bot's plans need heavy surgery every time, the input is wrong rather than the bot: usually the scheme of work file is vaguer than you think, and fixing the file fixes the output.

Second, keep the reuse note honest. The single most valuable line in the whole output is the one that points at a resource you already made, because most planning time is spent rebuilding things that exist. Once a month, spot-check three of those filenames actually resolve. A bot that starts confidently describing worksheets that do not exist has quietly become a cost.

Keep the roster small while you do this. Routines are per bot, capped at fifty per bot, and the app keeps the twenty most recent run records per routine, so a sprawl of half-used bots is not a filing system you can tidy later. Two bots you read every week beat six you skim in September and abandon by half term.

When the planner starts costing you time, check these six things

A planning bot fails quietly. It keeps producing plans, they keep looking reasonable, and the saving disappears into the editing. Each fix below is in the inputs rather than the prompt.

Check the second row first. A bot that confidently describes worksheets you do not have costs more than it saves, and takes a term to notice.

Where this guide assumes an institution you may not have

Everything above assumes a school: a written policy, an approved-tools list, and somebody whose job includes answering data questions. Three sets of readers lack that, and the advice moves.

A private tutor has no policy to check and nobody to ask. That does not remove the question, it moves it onto you, and the honest version of this article for a tutor is stricter rather than looser, because no second pair of eyes catches a mistake. Work from descriptions, and leave the gradebook line where it is.

A university lecturer working with teaching assistants is in a different shape, because marking is already distributed and a moderation step already exists. That makes the drafting case marginally stronger, since somebody other than the drafter reads the outcome. It leaves the reading argument untouched, because moderation samples and a teacher's own reading does not.

A trainer in adult or corporate education has no exam board and different record systems. Establish which regime you are under before importing caution from a school setting, and equally before importing permission from one. The planning half of this article transfers cleanly. The assessment half is written about children and records that follow them.

One thing transfers everywhere, tutor and trainer included: the edit is the occasion on which somebody reads the work. Remove the reader and you have not saved time, you have changed what the exercise was for.

Keep reading: Self-Describing CLIs, Grok Bot and X, Grok Bot and Zoom.

Frequently Asked Questions

What can AI bots do for teachers safely?

The safe surface is everything outside the gradebook. A bot can assemble a next-day brief of rooms, cover, deadlines, and what needs printing, track which topics in your scheme of work are planned versus taught versus behind, generate starters, misconception probes, and exit tickets for a topic, watch exam board and guidance pages for changes and quote the changed clause, turn long subject talks into notes, and sort logistics email with drafts attached. None of those require a student name, a piece of student work, or a mark.

Can a bot grade student work?

It should not, and in many institutions it may not. A bot must not grade, rank, or determine an outcome for a student on its own, which covers marks, set placement, entry tier, predicted grades, and pass or fail decisions. Drafting feedback that a teacher reads and edits before it reaches the student is the defensible scope, and even that is a data question first, so it depends on what your institution's policy allows. The reason is accountability: an outcome belongs to a person who can be asked to justify it to a student, a parent, or a moderator.

Is it legal to put student work into an AI bot?

Treat it as a decision with rules attached rather than a technical choice. Student data is legally protected in most places, including FERPA in the United States and the GDPR in the EU and the UK, and a bot's cloud computer is a third-party system under any definition your school uses. This is not legal advice. The practical step is to check your institution's policy first, because almost every school and district has one, along with an approved-tools list and a data protection contact whose job is exactly this question.

Why does the teacher have to edit the feedback?

Because the edit is where you read the work, and reading the work was the assessment. Marking produces comments as a byproduct, but its real function is finding out what a class understood, and a draft that lets you skip that has automated the reading rather than the writing. The edit is also where specificity comes from, since a model working from a rubric produces the median observation that is true of any piece of work. And it is where you can defend the comment later, which matters when a parent asks.