I ditched Claude Code and OpenCode for Pi, and my coding workflow became predictable again

A developer shares why they switched from Claude Code and OpenCode to Pi, a minimal terminal-based LLM harness. Pi uses fewer than 1,000 tokens for its system prompt, leaving more context window for actual work. Unlike other harnesses that inject hidden content into prompts, Pi gives full transparency and control over what enters the model's context. It also supports self-modification through extensions and encourages structured markdown files for planning, making workflows more reproducible and predictable.