Prompting still lives inside the loop

It is tempting to think looping replaces prompt engineering. It does not. Looping sits on top of prompting, and it makes your prompting skills more valuable, not less. Here is why.

Every think step is a prompt. Each time the loop turns, the model is deciding its next action — and that decision is driven by a prompt: the goal, the instructions, the tools available, and everything observed so far. A single turn of an agent is, at heart, one well-structured prompt. Which means the fundamentals you already know — being specific, giving context, stating the goal clearly, defining what done looks like — are exactly what make each turn of the loop sharp.

Good prompting makes the loop converge; vague prompting makes it wander. An agent given a fuzzy goal ("improve this") will loop aimlessly, trying things, unsure when to stop. An agent given a crisp goal, clear constraints, and a definition of success ("make all tests pass without changing the public function names; stop when the suite is green") loops with purpose and finishes faster. The clearer each think step, the less the agent drifts — and drift is one of the main ways loops fail (Module 2).

The highest-leverage prompt in an agent is the instructions. The standing instructions — the agent role, its rules, what it must not do, and what counts as finished — shape every single turn. Time spent making those clear pays off on every loop iteration, not just once. This is prompt engineering applied to a system that will use your words hundreds of times.

The takeaway for a prompt engineer. You are not throwing away what you know. You are adding one layer: instead of crafting one great prompt for one great answer, you craft the goal, constraints, and stop condition that a loop will execute against many times. The better your prompt fundamentals, the better your agents behave.

Recap. Looping does not replace prompt engineering — it runs on it. Every think step in the loop is a prompt, so being specific, giving context, and defining done make each turn sharp. Clear goals and constraints make a loop converge and finish; vague ones make it wander. The agent instructions are the highest-leverage prompt because they shape every iteration. So your prompting skills carry straight into agentic AI, and they matter more there, because the system uses your words again and again.