Delegate Outcomes, Not Steps
Don't steer AI line by line. Describe the final state you want, let it plan the path and report trade-offs.
Two ways people talk to AI: one says "change line 15 to this, delete that paragraph at line 20"; the other says "make this module support non-blocking reads while keeping the interface unchanged." The first uses AI as a typewriter, the second as an engineer. In vibe coding, the second is right far more often.
After reading, you should be able to answer:
- Why is line-by-line steering both slow and bad?
- How do you turn a requirement into "describing a result" instead of "describing steps"?
- When delegated outcomes drift, what do you do?
The problem with steering steps
- Slow: you give the steps, AI follows them, but you're not faster than it—you just recited its work back to it
- Throws away its strength: AI's edge is having seen many patterns and picking a better approach; lock it into dead steps and you've turned it into a typewriter
- You already over-thought it: line-by-line steering means the plan is already fixed, and you've given up the chance for AI to offer a better solution
How to write a delegated outcome
Write "goal + constraints + done criteria," and leave the path to AI anthropic-sdlc:
| Goal | what final state I want |
|---|---|
| Constraints | what not to touch, which approach |
| Done criteria | what counts as right, verifiable |
| Report | trade-offs and plan differences, reported to me |
Give it the minimum information and let go of the planning. Ask it to walk you through "how it plans to do it and what the trade-offs are" before it starts, then you decide whether to proceed. This uses its exploration power while keeping your judgment.
When it drifts
Delegating outcomes isn't walking away. When it drifts:
- Ask why it chose this first, don't immediately correct—it may see a trade-off you didn't google-patterns
- Say the goal, not the steps: "this isn't what I want; I want state X"—don't rewrite its path for it
- Let it fix the errors itself: if the first fix fails, tell it the outcome and let it try again
Outcome is judged by acceptance
Delegation rests on the "done criteria" you defined earlier. So the landing point returns to: can the result be verified? If yes, delegate big. If no, go back and raise the standard first.
Delegation must sit on top of "results can be accepted," and "accept results on evidence" is exactly the next chapter: Review the AI's Output.