Giving instructions to a child is a lot like working with AI. Well, not all of it, just the weak models. You tell a kid which soda to buy, and they forget about the trash they were supposed to drop off on the way. Remind them about the trash, and the soda drops out of their context.
Weak models work and forget in exactly the same way. While testing Plainline, the project I'm currently building, I tried both simple models with low reasoning effort and frontier models for generating presentation plans and slides.
The difference is interesting, and it's a lot like the difference between an adult and a child. Imagine going to the store for a two-liter bottle of Coke Zero. Both an adult and a child can handle this easily. But if the store is out of Coke Zero, an adult will figure out what to do on their own, say, grab a Pepsi Zero instead. They act proactively. A child will forget that it had to be Zero, grab a regular Coke, and take a 0.5-liter bottle because there were no two-liter ones. Or they'll just freeze and buy nothing at all.
LLMs work the same way. A strong model will fill in what you meant and try to act "sensibly." A weak one will behave just like that child at the checkout: it remembers one thing and forgets another.
That doesn't mean weak models (or children) are useless. They handle fairly complex tasks well, as long as those tasks are described properly. They just need clearer instructions and as few parallel tasks as possible in a single call. For example, generating only the plan in one call and only the text based on that plan in another works better than doing both in one call.
Research shows this too. In the IFScale study, models were given up to 500 instructions at once. Even the best ones followed only 68% of them at 500, and the instructions near the top of the list were followed more often.
When I went through my skill with a weak executor in mind, the first thing I removed was phrases like "if you think it's necessary." Weak models struggle to make decisions, so they'll either always skip such a step or always do it without thinking. And there's no way to predict which one the model will choose. So I rewrote every condition as a rule that can be checked without deliberation. The result is an almost deterministic algorithm.
I also had to name everything precisely: which files to touch, which commands to run. And I added a "When to stop" section. For example: if this happens, don't work around the problem, stop and report it. A strong model finds this unnecessary, but a weak one has nothing to rely on without it.
And of course, no decisions made "in its head." If a model does some work, it has to write down the result or the decision somewhere. The record itself works both as a log and as proof that the model actually did something instead of skipping a step.
So the rule is the same for children and models: the less room the other side has in their head, the shorter and clearer the list should be, and the more it matters what comes first.