Lada Kesseler: I Trust AI Tests Less Than AI Code

The discussion centers on the evolving role of AI in software development, emphasizing the importance of refinement, validation, and structured workflows. AI is portrayed not as a one-step solution but as a collaborative tool that requires iterative feedback to produce high-quality outputs. Key practices include using AI for prototyping, applying ground rules to ensure honesty and productivity, and breaking down tasks into focused steps - particularly in test-driven development. The concept of AI as a "centrifuge" is introduced, where rapid iteration helps separate valuable insights from noise, enabling better decision-making and system design.

A major theme is the shift toward agentic coding, where developers guide AI agents through defined processes rather than writing code line by line. This involves setting clear "front matter" for AI skills to prevent overlap and misfires, using verifiers to enforce standards, and integrating AI with existing tools like linters and formatters for automation. Challenges such as maintaining code quality, managing complexity, and avoiding premature implementation are highlighted, along with strategies like behavior-driven testing, approval testing, and event sourcing. The conversation also explores future possibilities, including AI-assisted architecture and software factories, while stressing the continued need for human oversight, trust in building blocks, and cautious, incremental adoption.

• What if you implemented AI ground rules to enforce honesty and productivity in your solo development workflow?

  • Move: Define and embed a personal "honesty + productivity" rule set in your AI prompts (e.g., "Always flag inconsistencies, never assume, and reject flawed logic even if it pleases me"). Store this in your prompt library or agent config (e.g., ClaudeMD-style user-level rules).
  • Why Now?: Default AI behavior is overly agreeable and hides edge cases - this leads to silent failures in solo projects where no one else reviews the logic. With AI coding autonomy increasing, now is the time to harden your feedback loop.
  • Expected Upside: Fewer flawed implementations, earlier bug detection, and higher trust in AI-generated code - saving 5 - 10 hours/month in rework and debugging.

• What if you used AI as a centrifuge to rapidly prototype and refine a new feature before writing real code?

  • Move: Run a 30-minute "spin cycle" where you prompt AI to generate, critique, and iterate on a mock implementation (e.g., API design or UI flow) using back-and-forth refinement - then extract only the stable output for actual development.
  • Why Now?: You're likely wasting time building features that shift after feedback. The AI centrifuge lets you fail fast in simulation, not production, especially as agentic workflows make prototyping nearly cost-free.
  • Expected Upside: Reduce feature rework by 40% and accelerate time-to-prototype from days to hours, giving you faster market validation on new ideas.

• What if you built a personal verifier agent to automatically check code quality and best practices post-generation?

  • Move: Create a standalone AI-powered "verifier" script that runs after each AI-generated code block - checking for simplicity, test coverage, and adherence to your rules (e.g., "no hockey solutions", "TDD followed", "no duplicated slices").
  • Why Now?: As a solo operator, you lack peer review - so letting AI generate unchecked code risks technical debt accumulation. With plug-in systems and deterministic triggers now accessible, automation is within reach.
  • Expected Upside: Maintain production-grade code quality without manual overhead, enabling 2x faster iteration while reducing bugs and refactoring needs by at least 30%.

• Define and implement personalized ground rules in your AI workflows to enforce honesty and productivity, such as instructing AI to challenge assumptions and flag potential issues early.
• Break down coding tasks into small, iterative steps with explicit instructions (e.g., TDD cycles) rather than relying on one-off prompts to improve AI output quality and maintain control.
• Use AI to generate drafts for code, emails, or documentation, but apply a structured refinement process - review, correct, and re-prompt - to achieve high-quality, reliable results.
• Create and refine "front matter" for AI skills or prompts to ensure precise, non-overlapping triggers that guide AI behavior effectively in specific contexts or projects.
• Integrate AI into targeted parts of your development workflow - like linting, refactoring, or test generation - using deterministic triggers (e.g., code smells) while retaining human review for critical decisions.