Stop Prompting Claude. Start Loop Engineering.

Rethinking AI Interaction: The Shift from Prompting to Loop Engineering

Introduction to Loop Engineering

• Boris Cherny, creator of Claude code, emphasizes a paradigm shift in AI usage: moving from prompting to designing loops.

• Peter Steinberg echoes this sentiment, suggesting that users should focus on building loops for coding agents rather than traditional prompts.

Understanding Loops vs. Prompts

• A loop is defined as a prompt that runs continuously until a specific task is completed, contrasting with one-time prompts.

• Boris explains his current workflow: he no longer prompts Claude directly; instead, he writes loops that automate the prompting process.

When to Use Loops

• Four conditions determine when to create a loop:

• Repetitive Task: Is the task something that repeats?

• Clear Definition of Done: Is there a clear way to define when the task is complete?

• Affordability of Waste: Can you afford potential token waste during execution?

• Necessary Tools Availability: Does the loop have all tools required for task completion?

Building Successful Loops

Key Components of Effective Loops

• The first building block is the trigger, which initiates the loop. Three simple methods include:

• Using slash loop for local machine tasks.

• Using slash schedule for cloud-based automation at specified times.

• Custom orchestration skills for comprehensive control over loops.

Execution Skills and Their Importance

• Execution skills are saved instructions allowing Claude to perform tasks consistently without retyping prompts.

• Each execution skill must be battle-tested; they ensure reliability in achieving desired outcomes within loops.

Goals and Verification in Loop Development

Establishing Goals and Verification Methods

• Every loop requires both a goal (what needs to be accomplished) and verification (how success will be confirmed).

• Technical examples involve measurable goals like website performance metrics; non-technical tasks require creative verification strategies.

Bridging Abstract Tasks with Verifiable Outcomes

• For non-measurable tasks, establish clear checkpoints using existing skills (e.g., email drafting verification through multiple review skills).

Output Management and Memory in Loops

Importance of Output and Memory Tracking

• Every loop generates output but also needs memory tracking to avoid repeating past mistakes or wasting resources.

• Documenting results helps improve future iterations by learning from previous outputs.

Practical Steps for Building Your First Loop

Starting Small with Skill-driven Development

• Begin by identifying small, proven tasks suitable for looping based on earlier discussions about conditions and components.

Utilizing Prompts Effectively

• Use provided prompts tailored towards building effective loops while leveraging past session history for inspiration.

Training Mode and Verification Checkpoints

Implementing Training Mode in New Loops

• Initially set new loops to pause at each step for user approval before proceeding—this ensures accuracy before full automation begins.

Key Checkpoints in Non-measurable Tasks

• Break down complex goals into smaller segments where human verification can occur at critical decision points throughout the process.