Stop Prompting Claude. Start Loop Engineering.

Too Long; Didn't Watch — Summary

Stop writing individual prompts and start designing 'loops'—automated, iterative AI workflows that run until a specific goal is verified as complete.

Main Takeaways

• The Paradigm Shift: Top AI developers are moving away from manual prompting toward 'Loop Engineering,' where the human's job is to design the system that prompts the AI.
• Loop vs. Prompt: A prompt runs once; a loop runs repeatedly, evaluating its own progress against a 'definition of done' until the task is finished.
• Skill-Driven Development: Successful loops should be built upon 'battle-tested' skills (saved sets of instructions) that the AI already performs reliably.
• Verification is Key: Every loop requires a bridge between abstract goals (e.g., 'good code') and verifiable outcomes (e.g., 'approved by a review agent').
• Memory and Efficiency: Loops must document their history in a markdown file or repository to prevent repeating errors and wasting tokens.

Detailed Summary

Introduction (0:00 - 0:46)

Industry leaders like Boris Cherny (Claude Code) and Peter Steinberger (Open Claude) are shifting from prompting to writing loops. This transition represents a higher level of abstraction in AI interaction, focusing on designing systems rather than individual messages.

Part 1: What a Loop Actually Is (0:46 - 2:34)

A loop is a prompt that executes repeatedly until a goal is met. To determine if a task is suitable for a loop, it must pass a four-condition test:

• Repetition: Is the task recurring?
• Definition of Done: Can the result be quantified or verified?
• Token Budget: Can you afford the potential waste of multiple automated runs?
• Tooling: Does the AI have the necessary tools (e.g., web access) to implement and verify the work?

Part 2: The Building Blocks (2:34 - 8:38)

Block 1: The Trigger (2:38 - 4:01)

The trigger is what initiates the loop. Methods include using /loop for local intervals, /schedule for cloud-based execution, or custom 'Loop Orchestration Skills'—a single command that kicks off a pre-configured workflow.

Block 2: Execution Skills (4:01 - 5:15)

These are specialized, saved instructions that perform specific tasks within the loop. Using 'skill-driven development' ensures the loop relies on methods already proven to work in isolation.

Block 3: Goal and Verification (5:15 - 7:53)

Every loop needs a goal and a rule to confirm completion. For non-technical tasks, you must bridge the abstract to the verifiable by using 'Review Skills' or scoring systems (1-10). Using separate AI agents to verify the work of the primary agent reduces bias.

Block 4: Output and Memory (7:53 - 8:38)

Loops must produce a tangible output (code, documents, messages) and maintain a memory file. Without memory, the agent 'forgets' previous failures, leading to infinite loops and wasted resources.

Part 3: Build Your First Loop Today (9:16 - 12:14)

Start small by identifying a proven skill and applying the four-condition test. Use 'Loop Training Mode' for the first few runs—a guardrail that forces the AI to pause and ask for human approval at every step. This prevents 'token burning' and ensures the AI doesn't go off-course during non-measurable tasks.

Notable Quotes

"I don't prompt Claude anymore. My job is to write loops." — Boris Cherny

"The agent forgets, the repo doesn't." — Adadi Asmani

"You don't build a loop without battle-tested skills behind it." — Austin Marchese