Outcome-Based Instructions
With today's more capable models, it's almost always better to prompt them with what your goal is, rather than how to get there.
The US military, despite being a massive organization, remains remarkably adaptive at the small-unit level because of two words: Commander's Intent.
Many large militaries enforce discipline by requiring low-level leaders to follow their commanders' instructions to the letter. That's not how the US military works. Whenever a US commander briefs his subordinates, he goes through a detailed plan, but it always starts with the commander's intent.
"Deny the enemy access to the bridge." That's the commander's intent.
"Alpha Company, you protect the right flank. Bravo, the left flank. Charlie, you cover the main approach. Delta, you provide overwatch." That's the plan.
When the plan crumbles like three-day-old cornbread, and the commander can't issue new instructions, his subordinate leaders can still take action. They use their own judgment to adjust on the fly. The commander's intent is their north star.
The destination is clear. How they execute is up to them.
Prompting with Intent
Today's frontier-level AI models work best when treated the same way.
Earlier models worked best when given strict instructions and told not to deviate from them: the more constraints, the better. Not so with newer models. If you give them too many constraints, they will eschew more efficient solutions that they would have otherwise pursued.
It's far better to provide your agent with verifiable outcomes to work toward, and then let it find its own way.
Mike Wolfe
Trustworthy Subordinates
Now, the whole "commander's intent" approach only works if you actually trust your subordinates to make good decisions on their own.
That means this is an approach that should only be used with more capable models. In the Anthropic ecosystem, that generally means Opus or Fable. Sonnet can do it–but not without close supervision from a higher-level supervisor (a human or an Opus/Fable-class orchestrating agent). Haiku and other similar fast/low capability models should only operate with explicit instructions.
As overall AI capabilities continue to improve, you should expect this approach to become the norm in most future situations.
