Custom Prompts Guide
Write prompts that give the same good answer every time, save them, and share them across a class or a department.
Last updated — 25 August 2026
What goes into a prompt
Four parts. Most disappointing answers are missing the fourth.
The four parts
- Role: who the model is answering as.
- Task: exactly what you want done, in one sentence.
- Context: the level, the subject, the syllabus, the audience.
- Format: how the answer should be laid out, and how long it should be.
Why format is the one people forget
A model that has not been told what shape the output should take will choose one — and it will choose a different one tomorrow. Naming the format is what turns a good answer into a repeatable one, which is the whole point of saving a prompt.
Writing and saving one
The prompt worth saving is one you have already typed twice.
Step by step
- Start from something you have already asked for more than once.
- Write all four parts out in full, including the ones that feel obvious to you.
- Run it, then fix the prompt rather than the answer, and note what you changed.
- Run it once more on a different input, to check it is not tuned to a single case.
- Save it under a name that says what it produces, not which subject it came from.
- Set the temperature with it: low when the same input should give the same output.
Leave gaps to fill in
Write the parts that change as visible placeholders — [subject], [grade level], [chapter] — so the prompt is reusable and it is obvious at a glance what has to be swapped before it is run. A placeholder you can see is a placeholder you will not forget.
Templates to start from
Four prompts that work as written. Replace the bracketed parts and adjust the format line to taste.
Study guide creator
Role: study guide creator. Task: build a revision guide for [subject], [chapter]. Context: for a [grade level] learner following [syllabus]. Format: key concepts, one worked example each, then eight practice questions with the answers at the end.
Step-by-step problem solver
Role: step-by-step problem solver. Task: solve the [problem type] problem I give you. Context: show the level of working expected at [grade level]. Format: numbered steps, the reason for each step, then the answer and one alternative method.
Answer sheet marker
Role: examiner. Task: mark the attached handwritten answer against the key below. Context: [subject] at [grade level], mark scheme as given. Format: marks per part with the reason for each, then two lines of feedback written for the student to read.
Image explainer
Role: subject teacher. Task: explain the attached diagram to someone seeing it for the first time. Context: [subject] at [grade level]. Format: what it shows overall, then each labelled part in turn, in no more than 200 words.
Keeping a library
A prompt library is only useful if a colleague can find the right one in ten seconds.
Organising
- Group prompts by what they produce — explanations, question sets, marking, image work.
- Tag by subject, so the maths ones are findable without reading every prompt.
- Name them for the output: "Chapter revision sheet" beats "Prompt 3".
- Retire prompts a newer one has replaced, instead of leaving both in the list.
Sharing across a class or department
A prompt is only text, so it travels. Hand the wording to a colleague and they get exactly your output format. Across a department this is what makes marking consistent between teachers, because everyone is asking for the same thing in the same words rather than each improvising.
Keeping versions
- Keep the working version before you edit a prompt that already does its job.
- Note what a change was meant to fix, so a later reader knows why the wording is odd.
- Date the version you standardised on, if a whole department is using it.
Habits that improve results
Six rules that apply whichever model you are running.
The short list
- Change one thing per attempt, or you learn nothing from the result.
- Be specific about the outcome and about who will read it.
- Show an example of the format you want when describing it is harder than showing it.
- Break a large request into smaller ones the model can actually finish.
- Say what to leave out as well as what to include.
- Lower the temperature when the same question should give the same answer twice.