looped-builder

The looped-builder skill is a specialized scaffolding tool designed for engineers and developers handling large-scale coding projects, extensive refactors, or deep technical explorations that exceed the context window of a single AI session. By transforming broad high-level requirements into a structured, file-based roadmap, it enables persistent and iterative development workflows. The agent creates a comprehensive specification (spec.md) and a granular implementation plan (implementation_plan.md) with atomic steps, followed by a local bash loop (run.sh) that allows Claude to execute tasks incrementally. Each step focuses on a limited scope, such as finding definitions, refactoring specific functions, or updating individual files, ensuring that the model maintains focus and accuracy throughout long-running operations. It is intended for complex features where manual management of state and progress becomes error-prone.

• Automatically generates project scaffolding including spec.md, implementation_plan.md, prompt.md, and run.sh.

• Breaks down complex requirements into atomic, context-aware steps suitable for Claude's context window.

• Implements an iterative bash loop that allows for pauses, resumes, and progress tracking via checkboxes.

• Includes built-in support for gathering codebase context, analyzing project patterns, and verifying success criteria.

• Provides a systematic approach to exploration, implementation, integration, and testing phases.

• Manages complex feature development by ensuring clear documentation and state persistence across multiple execution cycles.

• Ideal for large refactors, implementing extensive features, or complex system migrations.

• Users should define the end goal clearly to help the agent generate high-quality, actionable steps.

• Review the generated implementation plan before running the loop to ensure the logic and sequence match project requirements.

• The execution loop maintains status in the implementation_plan.md file, allowing users to monitor progress in real-time.

• The skill is compatible with standard bash environments and requires minimal configuration beyond setting the project root and iteration limits.

• Each step is designed to touch 1-3 files maximum, preventing context fragmentation and maintaining high code quality.