Been using PI Coding Agent with local Qwen3.6 35b for a while now and its actually insane

A developer has been using the PI Coding Agent with a local Qwen3.6 35b a3b q4_k_xl model for real production projects and reports surprisingly strong results. The key innovation is a custom "plan-first" skill file that enforces a structured workflow: the AI must analyze the project directory, dependencies, and build system, then ask up to five clarifying questions in a single round before writing any code. Only after the user approves a generated TODO.md does the agent begin executing tasks one by one. The developer claims this approach prevents the AI from going off the rails, making assumptions, or skipping steps—common failure modes in AI coding agents.

The workflow is designed for production-grade coding tasks including new features, bug fixes, and refactors. It requires the agent to read package files, build scripts, and existing documentation silently before asking questions. Each task in the TODO.md must be small and independently verifiable, ordered by dependency. The developer shared the skill file publicly for others to try. This approach addresses a major pain point with AI coding agents: their tendency to hallucinate or make premature changes without understanding the full project context. By forcing a plan-first methodology, the agent produces more reliable and maintainable code, potentially making local AI coding assistants viable for serious development work.