Loop Engineering: The Shift From Prompting AI to Building Systems That Prompt Themselves

The most important shift in how we build with AI is not a better model or a longer prompt. It is a structural change in how we think about the job. At Y Combinator Startup School 2026, Anthropic engineer and Claude Code creator Boris Cherny said it plainly: you are not supposed to prompt Claude. You are supposed to build a system that prompts itself.

Claude Code

Why prompting hits a ceiling

The one-request-at-a-time habit is the default. You type, read, refine, repeat. The AI never moves unless you push it. For a single task that is fine. For anything that repeats, like daily intel, weekly analysis, or recurring content, the bottleneck is not the AI. It is you, walking it through every step.

What a loop actually is

A loop gives the AI a goal and a way to know when it is done. It plans, executes, checks its own output, fixes what is weak, and repeats until the bar is met. Five building blocks:

1. The Trigger. What starts it: a schedule, an event, or a command.
2. The Skill. A reusable instruction set saved as a file the loop reads every time.
3. Sub-Agents (Maker and Checker). The model that writes is too generous grading itself. A second, stricter agent catches what the first one talked itself into.
4. Connectors. The difference between "here is the fix" and opening the PR, linking the ticket, and pinging the channel when it is done.
5. The Verifier. The gate: a test, build, or linter that auto-rejects bad work. Without it the agent grades its own homework.

If you want the deeper architecture, I have written about how the loop sits between the harness and the graph in Harness, Loop, Graph: The Three Layers Behind Reliable AI Agents, and about what changes when one loop grows into many in From Loop Engineering to Graph Engineering.

The self-checking pattern

You can run a basic loop right now in any LLM:

GOAL, DO, VERIFY, DECIDE. Give the model a goal. Let it do the work. Then make it score its own output from 1 to 10 against each criterion you set. If every score is 8 or higher, stop. Otherwise it fixes the weakest score and runs again.

When it is worth building

Four tests:

1. The task repeats at least weekly.
2. Something can auto-reject bad output.
3. The agent can do the work end-to-end.
4. "Done" is measurable, not a matter of taste.

What this means

The companies getting real leverage from AI have stopped treating it as a smarter search box. They build loops: a daily competitor brief by 7am, a customer signal engine that reads every call and surfaces growing patterns, a content pipeline that turns one voice memo into six platform posts.

The prompts still matter. But they are no longer the product. The loop is.

About the Author

I'm Brian Marvin, an AI-native Fractional CTO with 30 years in technical leadership. At empowered.guru, I help startups build MVPs, shape roadmaps, and turn AI from a chat window into managed, self-checking systems that scale.