Skip to main content

Guided Coding: Staying in Control While Coding with AI Agents

· 3 min read
Kenny Pflug
Kenny Pflug
Engineering Manager at TELIS/GWVS, Conference Speaker, Consultant

I recently joined Johannes Rabauer on his live stream to present Guided Coding — an approach I developed for writing enterprise-grade software with coding agents while maintaining full control over the source code. In this post, I want to summarize the most important points from that session. For the full details, check out the Guided Coding documentation.

You can find all the AI plans, AGENTS.md files, and the complete codebase at the Light.PortableResults repository. The full stream with Johannes is available on YouTube.

Guided Coding instead of Vibe Coding

Why Not Vibe Coding?​

Andrej Karpathy, co-founder of OpenAI, coined the term Vibe Coding in February 2025. The core idea: you tell the coding agent what to do, never go down to the source code level, always click "accept all changes", and the code grows beyond usual comprehension. You only validate by looking at the running app — never at the code itself.

While Vibe Coding is fun for personal projects and prototyping, it is not a viable approach for enterprise-grade software. You have no idea what the code is doing or how it's doing it, and technical debt accumulates rapidly. I wanted to formalize a counterpart — a methodology that lets you leverage the extraordinary productivity gains of coding agents while staying in control of the codebase.

That's what Guided Coding is about.

The Three Phases​

Guided Coding is structured around three phases: Plan, Implement, and Guide.

Guided Coding workflow diagram showing three phases: Planning, Implementation, and Guiding. Arrows indicate the flow between phases, with loops back to Planning for large issues and back to Implementation for small issues, ending with a PR.1. Planning Phase2. Implementation Phase3. Guiding Phase✓ PR
  • Planning Phase (~ 30–90 minutes) — Start a new conversation with your coding agent and iteratively create a plan together. Discuss the feature, commit the plan as a markdown file in git, and end with repeated sanity checks: "Would you add, change, or remove anything?"
  • Implementation Phase (~ 5–45 minutes) — Hand the agent a single plan and let it work autonomously, backed by feedback loops like compilers, automated tests, linters, and benchmarks. The agent fixes its own mistakes when these tools catch issues.
  • Guiding Phase (~ 1–6 hours) — The most important phase. Review all generated code thoroughly, evaluate the design, run the tests yourself, and do a manual smoke test. For small issues, iterate back to implementation. For larger problems, create new plans.

Beyond the Three Phases​

Guided Coding also covers several supporting practices:

  • Keep Your Rules Files Small — Start minimal, only add rules when agents repeatedly make the same mistake, and be aware of context rot.
  • Finding the Right Plan Size — Learn empirically what works for your model and codebase. You as the reviewer are the bottleneck, not the model.
  • Security: Be Vigilant — AI agents can be attacked through prompt injection. Review raw files, disable unnecessary MCP servers, and consider sandboxing.
  • Communication Is the New Superpower — The ability to clearly articulate what you want is becoming disproportionately important.

Practical Results​

I applied Guided Coding to build Light.PortableResults, an open-source .NET library implementing the result pattern with full serialization support. The codebase is roughly 25,000 lines of actual code as of February 26th, 2026, and more than 90% of it was written by AI agents. Depending on the feature, I estimate a speedup factor of 1.25x to 4x compared to writing everything manually.