The analogy

Give two helpers the same shopping list: fifteen items with conditions, "bread only if it's wholemeal, two cartons of milk if they're on offer, no frozen food if the queue is long". One comes back with everything right; the other forgets half the conditions and gets the wrong items. It's not bad will: it's how much each can hold together so many constraints at once.

And note one thing: both are better off with a written, ordered list, not one recited in a hurry on the way out the door. The AI is that helper. Some have a better "working memory" for rules, but they all perform better if the list is clear and in order.

How it really works

Following instructions is a skill that gets trained: the better-tuned models, and generally the more capable ones, manage to track more constraints at once without dropping any along the way. But there's a ceiling for anyone: too many instructions at once, or instructions that contradict each other, throw any model off, and it ends up satisfying some and ignoring others. The order and structure of the request help everyone: a numbered, isolated rule is easier to follow than one buried in a dense paragraph.

What you can do in practice

  • Number the instructions instead of slotting them into a block of text: the AI follows them one by one more easily.
  • When the task is complex, give one thing at a time, or break it into steps (the chained work).
  • Express the constraints that matter clearly and check that they don't contradict each other: two opposite rules cancel out.
  • Have it repeat the rules before executing: "repeat in your own words the rules you'll follow, then start". If it's missed one, you notice right away.

A common misconception

People think that if the AI doesn't follow everything, it's because it's "stupid". Often it's instead because you asked it for too much at once, or in a disordered and contradictory way. Even the best model loses pieces under an avalanche of piled-up constraints: it's not an intelligence test, it's a limit of how many things you can keep in focus together. Clear out and order the request, and the same "stupid" model follows everything.

Frequently asked questions

How do I get it to actually follow all the instructions?

Numbered list, no contradictions, and fewer things at a time when the task is heavy. If the rules are many, it's better to break into several requests instead of hoping the model holds it all together on the first go.

Which AI follows complex instructions better?

The ones better trained on instruction following, and generally the more capable models. But you notice the difference mostly on tasks loaded with constraints; on simple requests they almost all behave well. The way to find out is to try the same hard task on two tools.

Does the length of the request matter?

Yes. It's not the length in itself that's bad, but the pile-up: too many instructions crammed together step on each other. A long but ordered and numbered request works; a long, chaotic one doesn't.

The more detailed and long the request, the better it executes?

Up to a point. The right amount of detail helps, but beyond a threshold piled-up constraints and qualifications start to contradict each other and confuse the model, which lets some drop. Clarity and order beat quantity: better five clear instructions than twenty tangled ones. If you have to ask for a lot, break it up.