Goal, Context, Friction, Standard

Four things to put in a prompt: what you are trying to achieve, what the model needs to know about you, what it should make you do yourself, and what a good result would look like. The third one is the reason this exists. Every other published prompt framework asks what the model should be told. None of them asks what the person should be made to keep doing, and that omission is why a prompt can improve an output and cost the person the practice that produced it.

The four parts, and the one that is doing the work#

Goal. What you are actually trying to achieve, stated as the outcome and not the task. "Help me understand why this argument fails" rather than "summarise this".

Context. What it needs to know about you, your level and your material. Most bad output is a reasonable answer to a question about somebody else.

Friction. What should it make you do yourself, instead of doing it for you. Ask for the questions rather than the answers. Ask it to mark your attempt rather than replace it. Ask it to stop before the part you need to be able to defend.

Standard. What a good result would actually look like, said in advance, so you have something to judge the output against other than whether it sounds fluent.

A bad prompt does the hard part for you. A good prompt hands it back.

What the published frameworks ask for, and what they leave out#

Prompt frameworks are not scarce. The one with a peer-reviewed home is Leo Lo's CLEAR, published in the Journal of Academic Librarianship in 2023 and taught in university libraries: Concise, Logical, Explicit, Adaptive, Reflective. Google's TCREI runs Task, Context, References, Evaluate, Iterate. CO-STAR adds Style, Tone and Audience. Role-Task-Format is the short version most people actually use.

Read them next to each other and they agree on more than they disagree. Be brief. Be specific. Give it your material. Say what format you want. Look at what came back and go again. That is a good list, and if you are optimising an output it is the better list.

Every element in every one of them is an instruction to the model. Concise, Logical and Explicit describe the prompt. Adaptive and Reflective describe iterating on the prompt. Task, Context, References and Format describe the job. Not one of them contains a slot for what the person should be prevented from outsourcing, because output quality is the thing they are all built to improve, and on that measure holding something back is a cost.

Lo does not claim otherwise. CLEAR is offered as a framework proposal with no trial behind it, and its author presents it as a teaching aid for information literacy rather than as a tested intervention. The same is true of the others. It is true of this one.

Why a fourth slot for friction is not a preference#

The case for Friction is not that struggle is character-building. It is that the evidence on withdrawal is consistent and the evidence on output quality is beside the point.

Nearly a thousand school students were split between unrestricted GPT-4, a hints-only tutor and no tool. Grades rose 48 per cent with unrestricted access and 127 per cent with the tutor. When the tool was removed, the unrestricted group scored 17 per cent below students who had never had it, and the tutor group kept most of their gain. The difference between those two arms is a prompt-shaped difference: one answered, the other made the student produce something first.

In a study of 78 novice programmers, both AI groups beat the manual control on getting code working and were indistinguishable from each other until the tool was cut off, at which point unrestricted users failed a maintenance task at 77 per cent against 39 for a scaffolded group. And in randomised trials with 1,222 people, the withdrawal effect appeared after roughly ten minutes of interaction, showing up as reduced persistence rather than lost knowledge.

None of those studies tested this framework. What they establish is narrower and enough: the difference between an interaction that leaves capability behind and one that does not is a property of how the request is made, available at the moment of typing.

Friction in practice, and when to leave it out#

Concrete versions, because the abstract instruction is easy to nod at and hard to use.

Ask for the five questions you should be able to answer before reading the summary. Ask it to mark your draft against a rubric rather than rewrite it. Ask for the counter-argument to your position without the position restated. Ask it to stop at the outline. Ask it to give you the search terms and not the reading list. Ask it to interview you about the material for ten minutes and tell you where you were vague.

Friction belongs where the capability is one you need to keep. It does not belong on formatting, file conversion, transcription or anything you have no ambition to be good at, and adding it there is a way of making work slower without making anyone better. The test is the one on keep, share, hand over: if you could not defend the output in a room, it needed friction.

What this framework has not been shown to do#

No trial has tested Goal, Context, Friction, Standard against CLEAR, against TCREI or against no framework at all. Nobody has measured whether people who use it retain more, produce better work or notice more errors. The mechanism it is built on is evidenced; the framework itself is not, and it sits in exactly the position Hamilton and colleagues describe for SAMR: adopted through practice, absent from the literature, and popular for reasons that have nothing to do with whether it works.

It is also not the best-specified of these frameworks. CLEAR is tighter, TCREI is easier to teach, and both have more thought behind the mechanics of the instruction itself. If you want an output improved, use one of those. The claim here is narrower: they optimise the answer, and this one asks a question about the person that none of them asks.

Key sources

• Lo, L. S. (2023). The CLEAR path: A framework for enhancing information literacy through prompt engineering. The Journal of Academic Librarianship, 49(4).
• Bastani, H. et al. (2025). Generative AI Without Guardrails Can Harm Learning. PNAS, 122(26).
• Sankaranarayanan, S. (2026). Mitigating "Epistemic Debt" in Generative AI-Scaffolded Novice Programming. arXiv:2602.20206.
• Liu, G. et al. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance. arXiv:2604.04721.
• Hamilton, E. R., Rosenberg, J. M. and Akcaoglu, M. (2016). The SAMR Model: A Critical Review and Suggestions for Its Use. TechTrends, 60(5).

Related SuperSkills research#

The frameworks this one sits inside: think, AI, think and the four levels. On what to hand over at all, keep, share, hand over and the delegation boundary map. On the mechanism, desirable difficulty and productive struggle. On why prompting alone is weak advice, learn to prompt.

About this framework#

Goal, Context, Friction, Standard is used by Rahim Hirji in teaching and in the Mastering AI deck, most recently in September 2026. No claim of first use is made. No dated first publication exists for it, and a search of the Box of Amazing archive found none, so it is anchored to SuperSkills (Kogan Page, 2026) and to the deck rather than to a moment of coining. CLEAR belongs to Leo Lo, TCREI to Google, and CO-STAR and Role-Task-Format are in general circulation. The reading offered here, that published prompt frameworks optimise the output and none of them protects the person, is an interpretation by Rahim Hirji and is marked as an interpretation and not a finding.