Skip to main content
Flexibility is the skill — The goal isn’t to always operate at Level 4. It’s to match the level to the stakes. A throwaway script doesn’t need interactive code reviews. A payments refactor does.

​
The four levels

Vibe coding

Let AI go wild. No version control, no review, no guardrails. You accept whatever it produces and move on.When it’s appropriate: Prototypes, throwaway scripts, learning experiments, anything you’ll delete tomorrow.When it’s not: Anything that ships to users, touches production data, or lives longer than a week.

Agentic coding with discipline

You’re using AI to write code, but with basic engineering hygiene:
  • Git for version control — commit before and after delegating
  • File-level understanding of what changed and why
  • Basic security awareness (no secrets in code, no open permissions)
  • Sufficient test coverage to catch regressions
This is where most engineers should start. It’s fast and the safety floor is reasonable.

Agentic software engineering

The AI is doing more, but you’ve built guardrails around it:
  • Pre-commit hooks enforce formatting, linting, and secret scanning
  • CI pipelines catch what local checks miss
  • Function-level understanding — you know what every function does, even if you didn’t write it
  • Thorough testing — unit, integration, edge cases
  • One-shot AI code reviews — ask Claude to review the diff before you merge

High-quality software engineering

Maximum quality. You understand every line, and you use AI to deepen that understanding:
  • Line-by-line understanding of all changes
  • Self-reflection loops — “Are you sure about this approach? What about X?” Forces the model to catch its own mistakes before you have to
  • Interactive AI code reviews — back-and-forth dialogue about design decisions, not just a thumbs-up
  • Advanced research — use Claude to explore alternatives, read docs, compare approaches before committing to one
  • AI-powered quality control — background testing sessions, Playwright for web apps, automated regression checks

​
At a glance

Level 1: Vibe coding

Quality: Low · Speed: LightningTesting: Bare minimum · Reviews: None

Level 2: Agentic coding with discipline

Quality: Okay · Speed: FasterTesting: Sufficient · Reviews: Brief

Level 3: Agentic software engineering

Quality: Medium · Speed: FastTesting: Thorough · Reviews: One-shot

Level 4: High-quality SE

Quality: High · Speed: DecentTesting: Rock solid · Reviews: Interactive
More tokens ≠ more slop — A common misconception: spending more tokens means more AI-generated slop. That’s only true if the tokens go toward generating more code. When tokens go toward research, self-reflection, and deeper understanding, quality goes up — not down. Level 4 uses more tokens than Level 1. But the tokens buy verification, not volume.
← Prev: Agentic Engineering · Next: Robots-First →