project-development

Run a manual prototype by pasting one representative input into the model and evaluating output quality. Proceed if the task involves synthesis, subjective judgment, or natural language output; stop if it requires precise computation, real-time latency, or perfect accuracy.

Use five discrete stages: acquire, prepare, process, parse, render. Only the process stage calls the LLM, so the other deterministic stages can be debugged and re-run independently with file-system-based caching.

Default to single-agent pipelines for batch processing of independent items. Escalate to multi-agent only when you need parallel exploration, the task exceeds one context window, or specialized sub-agents demonstrably improve benchmark quality.

Multiply item count by tokens per item by price per token, then add a 20-30 percent buffer for retries and failures. The included pipeline template has an estimate command that projects input tokens, output tokens, and dollar cost from prepared prompts.

LLMs do not follow format instructions perfectly, so prompts need explicit section markers, format examples, and a statement that output will be parsed programmatically. Build parsers with flexible regex, sensible defaults, and error logging instead of strict matching.