- Loop engineering extends beyond single AI prompts.
- AI now iterates, verifies and refines autonomously.
- Review loops improve quality and reduce errors.
- Design AI workflows, not just better prompts.
For three years, prompt engineering defined AI interaction, with endless advice on crafting perfect questions for better results. While still useful for simple tasks, this is no longer the whole story.
Leading engineers now view prompting as just the starting point. Instead of single instructions, they are designing loop engineering workflows where AI repeatedly reviews, improves, and verifies its own output until a goal is met.
Nvidia CEO Jensen Huang said, “The future of programming is describing the problem, not writing the code.” echoing similar view, Boris Cherny, head of Claude Code at Anthropic said:
“I don’t prompt Claude anymore. My job is to write loops.” He said it on stage in June, and it’s been circulating through developer communities since.
The verdict? Prompting isn’t obsolete, but for complex work, iterative self-refinement by AI now outperforms expecting perfection from a single instruction.
Why One Prompt Often Isn’t Enough
Most people have already hit the limits of single-shot prompting. While AI can generate a usable draft instantly, the real value comes from the inevitable tweaks: adjusting tone, tightening sentences, or refining the opening.
These follow-ups aren’t just casual questions; they form a feedback loop where each response fuels the next improvement.
This trend shifts AI from a static search engine into a collaborator that revises work based on input. This pattern is universal across professions: journalists sharpen headlines, developers debug iteratively, and designers iterate on concepts.
The first answer is rarely the final one, and AI consistently performs better when allowed to refine its own work step-by-step.
Prompting vs. Looping
| Prompting | Looping |
|---|---|
| One instruction | Multiple rounds of refinement |
| Fastest results | Better final quality |
| Human guides every step | AI improves output through repeated iterations |
| Ideal for simple tasks | Better suited to complex work |
What makes looping attractive isn’t that it magically makes AI smarter. It creates opportunities for the model to identify weaknesses, correct mistakes and produce a stronger result before the work reaches the user.
Loop Engineering Is Something Different
While most users still manually manage AI loops writing follow-ups and judging results engineers like Cherny are pushing further toward autonomous agents.
Instead of waiting for human prompts, these agents can execute full cycles independently: a coding assistant might write code, run tests, fix failures, and rerun checks until success, only notifying the developer upon completion or when human judgment is truly needed.
Anthropic is already embedding this with tools like ‘/loop’, ‘/goal’, and ‘/schedule’ in Claude Code, letting users define success conditions and let the system handle the repetition.
As Google’s Addy Osmani notes, this marks a shift from prompt engineering to loop engineering: developers are no longer just asking better questions, but designing systems that intelligently decide when to continue, when to stop, and when to ask for help.
Manual Loops vs. Autonomous Loops
| Manual Loop | Autonomous Loop |
|---|---|
| User writes every follow-up prompt | AI determines the next step automatically |
| Human reviews every revision | Separate AI models can verify results |
| Takes minutes | Can run for hours or days |
| Common for writing and research | Common for software development and automation |
Every Loop Needs a Way to Stop
Granting AI autonomy introduces significant risks: a single mistake can be indefinitely repeated, generating dozens of incorrect outputs before anyone notices.
To counter this, Anthropic recommends using “one model to generate solutions and another to evaluate them,” reducing errors before outputs reach users. One model executes the task while another independently verifies the result. If the check fails, the loop continues until the issue is resolved or a limit is reached.
Additionally, Anthropic says CLAUDE.md files give Claude “persistent project memory so it doesn’t have to relearn context every session.” Without this, long-running workflows risk wasting time and resources by endlessly re-solving the same problems.
Before You Build an AI Loop
| Ask yourself | Why it matters |
|---|---|
| What does success look like? | Prevents endless loops |
| Who verifies the result? | Reduces hallucinations and errors |
| How much can this task cost? | Controls token spending |
| Should the AI remember previous work? | Avoids repeating mistakes |
| When should a human step in? | Keeps automation under control |
Prompting Isn’t Going Away. It’s Becoming the First Step.
Prompt engineering remains essential for daily tasks like emails, summaries, and translations, where a solid prompt plus a few refinements is usually enough. Loop engineering, however, shines for repetitive, predictable work like monitoring projects or analyzing feedback, allowing AI to operate autonomously without constant supervision.
The shift is philosophical. As Addy Osmani put it, “The job is no longer writing better prompts. It’s designing better systems.” Instead of judging a single response, users now define the desired outcome and let the AI determine the necessary steps.
Prompting isn’t disappearing; it’s becoming the first step in a larger workflow. The next valuable skill isn’t writing the perfect prompt, but designing the perfect workflow.
