# Custom instructions for AI—practical use cases *Written 10 September 2026.* The first thing I say to students in courses on digital tools for the music classroom is: never use an app in its default settings. They are almost always 90 bpm, C major, 4/4. None of that is neutral. Somebody chose those numbers, informed by a particular musical tradition, and if you do not change them, that person is quietly teaching alongside you. The same approach is relevant also for other digital tools. Every popular AI assistant has an equivalent screen. It is usually called *custom instructions*, *personalisation*, or *saved preferences*, it takes about a minute to find, and it is free on every plan I have looked at. Have you opened it? Most people, sadly, never do. I put the material below to the AI workshop for teachers at the University of Akureyri in February 2026, run with Kristian Guttesen and Magnús Smári Smárason. ## Why an empty settings box is a meaningful (and usually wrong) decision Two things need saying before the practical part: The first is that you are responsible for what you hand in, submit, publish or teach, whatever produced the draft. Verifying every source, catching the invented ones, and not letting a hallucinated sentence into your own text is *your* job and cannot be delegated. My own university's guidance for teachers says the same in its own way: anything made with the help of gervigreind needs review, adaptation and responsibility from the teacher, and none of it should be trusted without checking.[^1] I would put one thing more sharply than the guidance does—a model is not a source, and it does not get cited as one. I mention it because the rest of this post is about tools, and tools are where some people quietly stop applying it. The second is that this is getting harder to opt out of. AI is being folded into search, into office software, into the learning management system, into the phone keyboard. You may opt out of some of it, and whether (and when) I think you should, is a different discussion. Regardless, as it becomes a component rather than a destination, it becomes invisible, and what is invisible is less likely to be an area of agency. So—the time is now! Learn the basics, while this is still relatively easy to do. If you never open the settings, you work inside whatever defaults the company shipping the tool preferred. That is a choice being made on your behalf, by people whose interests are likely not identical to yours (wondering when 'social' medial stopped being social? Well...). For an educator I think this is not optional in the way it might be for someone else. Knowing your way around a library is part of the job. So is reading a news source critically. AI literacy now sits in the same place, and it sits there whether or not you use the tools yourself, because your students do. You cannot advise on something you have declined to understand. ## One: give it a hierarchy of sources An AI assistant with web access can be, on specific occasions, a better search instrument than a traditional search engine used the ordinary way. It reads more, it reads faster, and it will hold several sources against each other. This is a real advantage. Nut, left to its own priorities it will reach for whatever is nearest and most quotable, which means *the biases already in the material get repeated back to you* with the confidence of a summary. Telling it *what counts as a good source* fixes a surprising amount of that. Here is a stripped-down version of the hierarchy I run: <pre style="font-size:12.5px; line-height:1.45; font-family:var(--font-monospace),'SFMono-Regular',Menlo,monospace; background:var(--background-secondary); color:var(--text-normal); padding:0.9em 1.1em; margin:1.6em 0; border-radius:6px; overflow-x:auto; white-space:pre; tab-size:2;">SOURCES AND VERIFICATION Before answering any question about the present state of things—prices, specifications, versions, availability, who currently holds a role, recent events—search the web first. Do not answer from memory. If you cannot search, say so and mark the answer UNVERIFIED. Prefer sources in this order: 1. Peer-reviewed journal articles. Use PubMed, Google Scholar, or a subject repository. Prefer the most recent systematic review or meta-analysis where sources conflict. 2. Established news agencies and official documents from the responsible body. 3. Anything else, named as such. Search in the local language of the country or community concerned, not only in English. If no peer-reviewed source exists, say so plainly and use the most credible source you can find. If evidence is limited, mixed, or rests on a single study, say that too. If a page will not load, is paywalled, or blocks you, say so explicitly. Distinguish between "could not access X because Y", "accessed X but it did not contain this", and "searched and found nothing". Never write as though you had read something you did not retrieve. Flag anything shaky as you go, using: MISINFORMATION (my question contains a myth), DISPUTED (sources genuinely disagree), UNCERTAIN (the evidence is thin or conflicting), UNVERIFIED (no confirmed source). End every answer containing a factual claim with an APA-7 reference list with working links. A missing reference list means there were no factual claims, not that you decided to skip it.</pre> The reference list at the end does two jobs. The obvious job is that it lets *you* check. You can click the links, read each one to see whether it says what the answer claims, and find out quickly which parts of a fluent paragraph were built on nothing (if it happens to be the case). Note the failure mode you are checking for: a fabricated reference is not a garbled one. It is a plausible title attached to a real author, with a DOI that resolves to nothing. The less obvious job is that the demand changes *what the model does before it writes*. A question answered from a model's own weights is recall, and recall is where invention happens. A question that has to arrive with links attached pushes the system towards retrieving something and reading it. You are not making the model more accurate by asking. You are making it work in a mode where its accuracy is checkable, and you are removing the option of a confident answer with nothing underneath. Keep the two apart when you explain this to students, because the difference is the whole point. *Listed citations do not make an answer true.* They make it possible to verify. ## Two: make it show its working *Translation* is where the frontier models have become quietly very good, at least for everyday communication and low-stakes text (they can still make rather unpredictable mistakes!). If you have not started using AI for translation, it is worth trying. If you have, the difference between a serviceable translation and a good one is almost entirely in the instructions. The same principle applies as above, in a different shape. Ask for the output and you get the output. Ask for the *steps*, visibly, and the work improves—partly because the model has to commit to a first version and then argue with it, and partly because you can see where it is guessing. These are my instructions for translating from English into Icelandic, somewhat simplified. They are set up for Gemini and work as they are; adapt the dictionaries and the register lines to your own needs. You can try it with any language(s). <pre style="font-size:12.5px; line-height:1.45; font-family:var(--font-monospace),'SFMono-Regular',Menlo,monospace; background:var(--background-secondary); color:var(--text-normal); padding:0.9em 1.1em; margin:1.6em 0; border-radius:6px; overflow-x:auto; white-space:pre; tab-size:2;">WHENEVER ASKED TO TRANSLATE BETWEEN ENGLISH AND ICELANDIC You translate into Icelandic for an academic and educational context. PRIORITIES Prioritise natural Icelandic phrasing, flow and idiom over literal rendering. Avoid anglicisms. Use the gender-neutral phrasing standard in contemporary Icelandic higher education. Match the register to the reader: clear, warm and supportive for students; formal and precise for colleagues and for academic submission. Prefer descriptive prepositional phrases to heavy compound nouns when referring to parts of a document. Use natural passive phrasing for document distribution. Keep items and quantities in the nominative after a colon in lists. Default sign-off: Bestu kveðjur, VERIFICATION Check grammar, vocabulary and contemporary usage against Málið.is (Árnastofnun) at https://malid.is and Tímarit.is at https://timarit.is. For educational terminology, give priority to the official terms used by Icelandic universities and to Orðasafn í menntunarfræðum. WORKFLOW Pass 1 – translate for natural syntax, register and target terminology. Output the full translated text. Pass 2 – audit that draft for gender agreement, inflection, cadence and terminology, checking against the sources above. Output the corrected version. Where sources conflict or you are unsure, say so rather than choosing silently. OUTPUT FORMAT – mandatory for every translation ✍🏻 First pass [first-pass translation] --- 🤔 Yfirlit yfir leiðréttingar (Pass 2 audit log) - [list the specific corrections made in Pass 2: grammar, inflection, register, terminology. Always disclose these. If nothing changed, say nothing changed.] --- ✅ Final translation [final Icelandic text] --- 🎓 Orðalykill / Glossary of key terms - [itemised list of specialised terms used, cross-referenced against the official terminology sources]</pre> All three parts earn their place. *The audit log* is where you find out what the model was unsure about, which is usually the sentence you should look at yourself. *The dictionary check* is what stops it inventing a plausible Icelandic compound that no one has ever written down. And *the glossary at the end* is the part that compounds: after a dozen translations you have your own terminology list, built out of your own documents, and you can start correcting it. ![[Gemini [email protected]|698]] *Gemini with default settings* ![[Gemini [email protected]]] *Gemini with custom prompt* The text in both examples comes from the chapter on *lykilhæfni*, the key competences that run across every subject, in the Icelandic national curriculum for compulsory school.[^2] ![[Official [email protected]]] *The official translation* ## Three: the questions where the “quickest” answer is wrong Here are two questions I have been putting to AI assistants in workshops (I have not come up with them myself, I found them on AI-related forums). Try them on whatever you use. > I need to charge my car. The charger is 100 metres away. Should I drive or walk? > I have a metal cup. The bottom is missing and the top is sealed shut. How can I use this cup? ![[[email protected]|700]] ![[CleanShot 2026-09-10 at [email protected]]] Before you ask an AI what is wrong with these, *sit with them yourself for a moment*. That part is not decoration. The value of the exercise is not in collecting two puzzles; it is in learning to recognise the shape of a question that is going to go badly, so that you write your own instruction the next time you meet one. Copying my blocks is the easy half. So: what is the problem? In the first, "drive or walk" is offered as though the two were interchangeable ways of covering a hundred metres. They are not, because the car is not your transport in this question. It is the thing that has to arrive. Walking gets *you* to the charger and leaves the car exactly where it was, which fails the only condition that mattered. This is not a question of which option is more efficient. One of them does not do the task. In the second, the word *cup* arrives carrying an orientation with it. A vessel with no bottom and a sealed top is useless the way up you are imagining it. Turn it over. The sealed top is now a base and the missing bottom is now a rim, and you have a cup. Obvious, right? Well, depends on the reasoning applied. Answers that offer you a vase, a napkin ring or a sculpture have accepted the orientation as given and gone looking for a clever way around it, missing the simple practical solution. Both failures are in fact the same failure. The system *matched the question to the shape of a question it has seen many times*—a distance question, an unusual-uses-for-an-object question—and answered that one instead of this one. It is a reasonable description to say the answer gets pulled towards what is most commonly said rather than towards what makes sense here—these systems reproduce and amplify the dominant patterns in the material they were trained on, which is the well-documented part of the story (Bender et al., 2021; Birhane, 2021)[^3]—and the pull is strongest exactly where you would least want it: specific, unfamiliar problems that turn on cause and effect. Why does it matter to you, you wonder? Imagine structuring a complicated itinerary with the help of AI. You might need to charge a car... Or get on a train in time, considering a reasonable number of factors which might make it (im)possible. The number of everyday life cases when biases like this can come into play is greater than one might assume at first. I would put one caution next to that, having watched these two examples for a while. The models are getting better at precisely this, and both questions are now answered correctly by some assistants that failed them a year ago (not by all though, as of today, which I find somewhat surprising!). Do not expect my examples to keep working some time from now. *What will keep working is the underlying habit*: a question whose surface strongly resembles a familiar one is a question to check. The instruction I use against this is the one part of my setup I will give you in full, unedited, because taking a piece out of it breaks it: <pre style="font-size:12.5px; line-height:1.45; font-family:var(--font-monospace),'SFMono-Regular',Menlo,monospace; background:var(--background-secondary); color:var(--text-normal); padding:0.9em 1.1em; margin:1.6em 0; border-radius:6px; overflow-x:auto; white-space:pre; tab-size:2;">🧩 Practical questions. VERY IMPORTANT—choose the action that directly changes the state of the object that needs to be acted upon. Before responding to any question involving a choice between options, work through these steps silently: 1. State the goal—what outcome needs to be achieved? 2. State the dependencies—what does each option require to succeed? 3. Eliminate incoherent options—does any option fail a necessary precondition? If so, eliminate and explain. 4. 🔄 Object check. When a physical object is involved, check whether reorienting, reversing, or repositioning it resolves the problem trivially before generating creative workarounds. 5. Only then choose, reasoning further among surviving options. Be especially alert to questions where one option appears obviously correct by a surface heuristic (distance, cost, effort) but silently violates a dependency stated elsewhere (e.g. the object cannot be separated from the agent acting on it). If no option survives, state this explicitly and explain which precondition each fails.</pre> The two puzzles are handled by two different lines of it, which is why both are in there. The car is caught at step three: walking fails a precondition of the goal, so it is eliminated before any comparison of effort happens. The cup is caught at step four, which exists solely because I got tired of being offered napkin rings. ![[CleanShot 2026-09-10 at [email protected]]] *Response with custom instructions (from Claude, so a different pair of shoes, anyway, but works similarly with other models).* ## Where this stops The blocks above are free to take and adapt. But they are intended to be a starting position rather than a solution, and I would rather you finished this post with a method than with a clipboard. Frankly, the limit on all of it is you. Not the model, not the plan you are on, not which company's assistant you happen to have. The tool amplifies whatever attention you bring to it and quietly substitutes its own defaults for whatever attention you withhold. Agency here is use it or lose it in a fairly literal sense: **the less you exercise, the less you are offered, and there is no rule saying it stays available to people who have stopped asking for it.** Whether you use these tools, and how much, is properly your decision, and the practical and ethical arguments on either side are worth having out. Not knowing about how they work is a different matter. That is not a neutral position, and for an educator it is not a defensible one. *See also:* [[Futures literacy]] [^1]: Setberg. (n.d.). *Ábyrg notkun í kennslu* [Responsible use in teaching]. Kennslumiðstöð og kennslusvið, Háskóli Íslands. https://setberg.hi.is/is/notkun-i-kennslu. The page states that content produced with the help of AI "þarf alltaf rýni, aðlögun og ábyrgð frá kennara" and warns against trusting its output blindly. The University's adopted AI policy adds that students and staff "bera ábyrgð á öllu efni sem þau leggja fram í eigin nafni": https://hi.is/haskolinn/gervigreindarstefna_haskola_islands [^2]: Miðstöð menntunar og skólaþjónustu. (2024). Kafli 18: Lykilhæfni. In *Aðalnámskrá grunnskóla: Almennur hluti*. https://www.adalnamskra.is/grunnskoli/kafli-18-lykilhaefni-2024 [^3]: Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In *Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency* (pp. 610–623). https://doi.org/10.1145/3442188.3445922 · Birhane, A. (2021). Algorithmic injustice: A relational ethics approach. *Patterns, 2*(2), 100205. https://doi.org/10.1016/j.patter.2021.100205