project-development

Run the task through the task-model fit tables: proceed if it involves synthesis across sources, subjective judgment with rubrics, or natural language output with error tolerance. Stop if it requires precise computation, real-time responses, or perfect accuracy. Always validate with a manual prototype first.

Use the five-stage architecture: acquire, prepare, process, parse, render. Keep only the process stage non-deterministic, store intermediate outputs as files per item for idempotency, and parallelize LLM calls with ThreadPoolExecutor. The included pipeline_template.py implements this pattern.

Default to single-agent pipelines for batch processing with independent items. Escalate to multi-agent only when parallel exploration is required, the task exceeds a single 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% buffer for retries and failures. The pipeline template includes an estimate stage that counts prompt characters and projects input and output token costs before processing.

LLMs do not follow format instructions perfectly, so brittle parsers break on minor variations. Use flexible regex patterns, provide sensible defaults for missing sections, log parse failures instead of crashing, and state in the prompt that output will be parsed programmatically.

Reduce tools when the data layer is well-documented, the model has sufficient reasoning capability, and scaffolding constrains rather than enables performance. Vercel's d0 agent improved from 80% to 100% success by cutting 17 tools down to bash and SQL execution.