project-development

Check task-model fit before writing code. Tasks suited to LLMs involve synthesis across sources, subjective judgment with rubrics, natural language output, and error tolerance. Avoid tasks requiring precise computation, real-time responses, or perfect accuracy. Always run one manual prototype through the model interface first.

Use a five-stage architecture: acquire, prepare, process, parse, render. Only the process stage calls the LLM; all other stages are deterministic. Store intermediate outputs as files per item so stages are idempotent, cacheable, and independently re-runnable.

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. Multi-agent costs roughly 15x more tokens than chat.

Multiply item count by tokens per item by price per token, then add a 20-30% buffer for retries and failures. Estimate input tokens from prompt length and output tokens from typical response length. Track actual costs during development to catch overruns early.

LLMs do not follow format instructions perfectly, so rigid parsers break on minor variations. Use flexible regex patterns that tolerate formatting differences, provide defaults for missing sections, and log parse failures for review instead of crashing.

Usually not. Vercel's d0 agent improved from 80% to 100% success by reducing from 17 tools to 2 primitives. Start minimal with file system access and add complexity only when production evidence proves it necessary, since scaffolding often constrains capable models.