Stop Blaming Models: Your Agent Prompts Are the Real Bottleneck

Theo details a systematic approach to improving AI coding agent outputs by refining prompt files rather than modifying model behavior. He spent 12+ hours rewriting his agents.md and CLAUDE.md files, creating specialized skills for PR management, file uploads, and HTML communication.

He describes a critical insight: the value isn't in copy-pasting his configurations but understanding the process of diagnosing agent failure modes through log analysis. He had multiple models audit his chat histories to identify common mistakes like Opus 5 killing running processes or agents filing excessive draft PRs.

The real tension lies between agent productivity and communication quality. His most impactful changes weren't technical—they focused on making agents better at describing problems, avoiding scope creep, and producing readable outputs. This is workflow design, not code optimization.

Engineering leaders should recognize this as an organizational pattern: investing in agent communication patterns and failure analysis yields bigger returns than chasing better models. The bottleneck isn't model capability but prompt infrastructure and workflow design.