project-development

This skill provides a rigorous methodology for building, scaling, and maintaining LLM-powered applications. Designed for software engineers and AI architects, it focuses on the operational challenges of LLM integration, specifically addressing the gap between initial prototyping and production-grade stability. The methodology emphasizes the use of staged pipelines to isolate non-deterministic LLM processes from deterministic data handling, ensuring that systems remain debuggable, cost-effective, and highly performant.

• Systematic Task-Model Fit Analysis: Procedures for deciding whether to use LLMs or traditional code, based on task characteristics like synthesis, judgment, computational precision, and latency requirements.

• Canonical Pipeline Architecture: Implementation of a structured process flow (acquire, prepare, process, parse, render) that separates complex reasoning steps from data retrieval and output formatting.

• Filesystem-as-State-Machine: A pattern that utilizes the file system to manage pipeline state, enabling idempotency, easy debugging, and simple caching of intermediate outputs.

• Structured Output Design: Techniques for generating parseable, schema-compliant LLM responses using section markers, format examples, and constrained values.

• Use this skill when initiating new agentic projects, planning batch processing pipelines, or troubleshooting unreliable LLM outputs.

• Focus on the 'Manual Prototype' step before automation; it serves as a baseline to identify knowledge gaps or failure modes in prompts.

• Maintain idempotency by using file-based storage, which allows for selective re-execution of expensive LLM steps without re-running deterministic data acquisition.

• Prioritize error-tolerant architectural design, as individual LLM failures should be handled via graceful parsing and logging rather than systemic crashes.

• Integrate with agent-assisted development tools to iterate on pipeline stages rapidly, ensuring that parsing regex and rendering logic are robust against non-deterministic model variance.