In a nutshell Property #4 of four: how far your instructions actually steer the model. The good news: a lot — a clear prompt steers the output almost like a wheel. The caveat: instructions steer, they don't guarantee. There are limits — contradictions, a safety priority, "almost, but not quite". Understanding those limits is how you get what you need reliably.
What steerability is
Steerability is the degree to which the form of your prompt changes what the model produces: tone, format, role, length, style, depth. Modern models are highly steerable: the same question, phrased differently, gives answers of completely different shape. What you steer in practice:
- role: "explain like a teacher to a first-grader" vs "like an expert to a colleague";
- format: table, list, JSON, paragraph, N points;
- tone: formal / friendly / strictly to the point;
- length: "in one sentence" / "in depth with examples";
- rules: "answer only from the given text".
Analogy. The model is a very capable but literal worker. Say a vague "make it nice" and you get its version of nice. Give a precise brief with a sample and you get near-exactly that. There's a wheel, but you must turn it deliberately.
The strong levers
- be specific: not "shorter", but "at most 3 sentences"
- examples (few-shot): 1–2 samples of what you want — the most powerful lever
- role and context: "you are an editor at a science journal"
- structure: instruction separate from data, an exact output format
- iterate: say what exactly was off — that's normal
Where the wheel slips
- contradictions: "in depth, but in one word"
- safety and principles outweigh your instruction
- "almost, but not quite": asked for 5 points, got 6
- too many rules at once — some get lost
- vagueness: the fuzzier the ask, the more it "fills in"
The limits of steerability
Instructions are strong influence, not absolute control:
- Contradictory directives. "In depth, but in one word" — the model picks one or produces a compromise. Check your prompt for internal conflicts.
- Safety priority. Part of the behaviour is set by training (see how generative AI works) and outweighs your instruction: requests for harm are refused, however you ask. Not "disobedience" — a built-in boundary.
- "Almost, but not quite." Sometimes it follows mostly, not perfectly: asked for no preamble, it adds an intro. The nature of token prediction (see next-token prediction) doesn't guarantee a strict format 100%.
- Too many rules. Twenty requirements in one prompt — some get lost. Split the task or set priorities.
Takeaways
- Steerability is high — but you own the clarity. A vague prompt → an answer "at the model's discretion".
- The strongest levers: specifics + examples + structure.
- Not everything bends: safety and principles override instructions; a strict format isn't guaranteed — check it.
- "It didn't obey" — first check your prompt (contradiction? vagueness? fell out of context?), then conclude it's a limit.
Source
Anthropic Academy course AI Capabilities and Limitations, section "Steerability".