project-development

This skill provides a systematic methodology for architecting and developing LLM-powered applications. It is designed for engineers, architects, and developers who need to evaluate whether a specific task is suitable for language model automation or if it requires traditional deterministic code. The skill focuses on reducing development waste by enforcing a rigorous 'manual prototype first' policy and utilizing a standardized pipeline structure to separate non-deterministic LLM processing from deterministic data handling.

• Task-Model Fit Analysis: Procedures for determining if a task involves synthesis, subjective judgment, or batch processing, and identifying 'stop' indicators like real-time requirements or exact math.

• Canonical Pipeline Architecture: Implements the acquire-prepare-process-parse-render pattern, enabling modular debugging, idempotency, and efficient cost control by isolating expensive LLM calls.

• File-System-as-State-Machine: Uses directory structures as a persistence layer to ensure pipeline state is human-readable and easily debuggable, facilitating selective re-execution.

• Structured Output Design: Best practices for prompt engineering that prioritize parseability, including section markers, format examples, and constraint validation.

• Agent-Assisted Development: Best practices for iterating on codebases with LLM-powered coding agents.

• Use this skill when starting a new LLM project, designing batch pipelines, or estimating the costs and timelines of an agentic system.

• The workflow requires a manual validation step where the user tests inputs against the LLM interface before investing in full-scale automation.

• Expected inputs include project requirements, task descriptions, or proposed architectural goals. Outputs generally consist of a structured pipeline blueprint, a go/no-go assessment for automation, and prompt design templates.

• Constraints: The skill is not suitable for high-precision, sub-second latency, or purely deterministic algorithmic tasks; it prioritizes iterative, scalable LLM workflows.