Why does AI follow some instructions perfectly and others loosely?
Some requests AI follows to the letter. Others are more of a suggestion.
You ask for a table with three columns. You get exactly three columns.
You ask for 200 words. You get 260.
You ask it not to mention the price. It writes, "I won't go into pricing here, but…"
Why the difference?
Made-up example.
Steerability: how easily you can change what the model does just by asking. A car with good steering goes where you point it. A shopping trolley with a wonky wheel sort of does.
Anthropic's course on AI's capabilities and limitations makes the point that steerability isn't the same for every kind of instruction. Some are tight, some are loose.
These tend to land well:
These are things it has practised constantly, and they're easy to check: either it's a table or it isn't.
These tend to slip:
Anthropic's own prompting advice: say what you want instead of what you don't.
| Instead of… | Try… |
|---|---|
| Don't use jargon. | Write for someone who's never worked in IT. |
| Don't make it too long. | Keep it to about five sentences. |
| Don't mention pricing. | Cover the features and the timeline only. |
| Don't be so formal. | Write it like a message to a colleague you like. |
"About 200 words" usually lands somewhere between 150 and 300.
You'll get closer with a shape it can see:
If the number really matters (a form with a 250-character limit, say), check it yourself or ask it to count with a tool.
Put the tight things in the instruction and check the loose ones yourself.
Rewrite your "don'ts" as "dos".
Give lengths as shapes: paragraphs, bullets or slides, not word counts.