AI can help you write software faster. It can also help you generate elegant-looking nonsense at industrial scale.

In this session, I’ll share the workflow I currently use to make AI-assisted development genuinely useful without letting it turn into slop.

Over the last years, I’ve tested different models, harnesses, loops, orchestrators, and agent setups, and gradually moved toward a style of working that is less about chasing autonomy and more about keeping context under control, asking better questions, and validating aggressively.

I’ll walk through the journey: why I standardized on OpenCode as my main harness, how Research – Plan – Implement became the backbone of my workflow, and how I scale the process up and down depending on the task instead of forcing one ritual onto everything. I’ll also cover a newer RPI v2 flow I’m exploring and automating.

A big part of the talk is about the difference between “AI helped write code” and “AI helped ship correct software.” I’ll show how I use plans as real handoff artifacts, when I split work into parallel streams, how I use orchestrators versus coding subagents, and why browser verification, tests, review, and concrete acceptance signals matter more than whichever model or harness is currently trending.

This is not a universal best-practice talk. It’s a practical field report from someone trying to become a more professional engineer with AI in the loop, not a faster producer of polished-looking mistakes.

Assumed Knowledge:

Attendees should already be familiar with normal software engineering practices such as version control, testing, code review, and working in a real codebase. Some prior exposure to AI coding tools like GitHub Copilot, OpenCode, Claude Code, or similar agent-style workflows will help, but deep experience is not required.

Practical Takeaways:

Attendees will leave with a concrete workflow for using AI without defaulting to vibe-driven development. They will understand when a task needs direct prompting, when it benefits from planning, and when it should be broken into research, design, structure, and parallel workstreams. They will also get practical ideas for orchestrator instructions, validation loops, browser-based verification, and using plans as operational artifacts instead of generated paperwork.

Out of Session Scope:

This session will not be a model benchmark shootout or a comprehensive comparison of every coding harness and agent framework. It will not teach attendees how to build their own autonomous coding platform from scratch, and it will not present a one-size-fits-all process for every team. The focus is on a practical anti-slop workflow for real software delivery, not full automation or theory for its own sake.