RTX Core Framework — Autonomous Agent System Skill

The RTX Core Framework is a field-tested master system-prompt blueprint that transforms a generic AI assistant into a highly autonomous, personalized execution system — designed to run across any agent that accepts a system prompt or SKILL.md file. Its signature feature is a 70% Romanized mother-tongue + 30% English language blend: you collaborate with your AI in your own language's phonetics (Hinglish-style) without mastering English, while code, terminal output, and schemas stay in pure English via a built-in token-deflation safety valve. On first contact, the agent runs a three-question First-Boot (mother tongue → agent name → how to address you), then injects a persona behind the name you choose — researched, not hardcoded. Every response follows a strict 3-step output protocol, and every task runs an Explore → Plan → Execute → Verify → Summarize loop with zero-tolerance internal review checklists. Nothing is declared "done" without attached artifacts: test counts, terminal transcripts, before/after numbers. Multi-agent work follows a commander/soldier doctrine (parallel read-only workers, one writer), and the repo ships a self-evaluation kit — 20 golden tasks plus a pass@k scorer — so you can verify compliance on your own models instead of trusting claims. Non-destructive actions run autonomously; high-risk ones pause for approval. All installation steps are explicit and opt-in — nothing runs in the background. MIT-licensed, by Prosenjit Paul (aka Prosen).