The final part of the divvy-forge series: wiring TrueForge deployment, discovering the real agent manifest API through trial and error, fixing coordinator prompt flow so subagents run silently, and watching the agent open a real GitHub PR with coordinator reasoning, a unified diff, and subagent findings.

How I designed the coordinator and subagent prompts for divvy-forge as versioned markdown files, embedded subagent instructions inline, defined strict output contracts, and used promptfoo evals to catch three real prompt bugs before a single live agent run.

A practical engineering guide to four core LLM prompting techniques—system prompts, few-shot examples, chain-of-thought reasoning, and structured output—covering real failure modes and production-ready patterns.