Inspired by Pi (pi-mono) and empirical findings from Harness Engineering (arXiv:2609.00006), POCLI replaces monolithic 15,000+ token prompts with role-scoped harnesses of 150–250 tokens.
Why Prompt Size Matters
Every token in a system prompt is re-processed on every conversational turn, multiplying API cost and Time-To-First-Token (TTFT)—especially on local LLMs (Ollama, llama.cpp) or free API tiers.
| Environment | System Prompt Overhead | Design Approach |
|---|---|---|
| Claude Code | ~25,000 tokens | Monolithic tool manuals and style rules |
| OpenCode / Codex | 10,000 – 15,000 tokens | Heavy multi-tool JSON schemas |
| Pi Agent | 500 – 900 tokens | Minimalist core guidelines |
| POCLI | 150 – 250 tokens | Ultra-concise, role-scoped harnesses |
Preserving 90–95% of Your Context Window
- Each agent receives only the tools required for its role (
Architect: ~140 tokens,Developer: ~105 tokens,QA: ~70 tokens,PM: ~130 tokens). - Total turn payload averages 600–1,200 tokens, yielding sub-second TTFT while leaving 90–95% of context capacity for source code, diffs, and RAG citations.
- Prompts reside cleanly in
behavior/system/*.mdand support user-level overrides in$HOME/.pocli/behavior/system/.
