Recap and Practice
Key takeaways
- Vibe coding is directing an AI to build software while you stay in charge of what and why — not blindly shipping code you can't explain.
- The rule that holds it together is understand what you ship: you don't have to type it, but you must be able to defend it and fix it.
- Your bottleneck moved from production (typing fast) to specification and review (describing clearly, judging critically).
- Live in the loop — intent → generate → review → refine — and never skip the review step, where the actual engineering happens.
- The model's confidence is not a signal; your verification is. Taste and judgment are the scarce, compounding skills.
Try it
Pick a small function you genuinely understand — something like "validate a password meets a policy" or "format a byte count as a human string." Write a precise for it (signature, rules, cases, a request for tests), generate it, then deliberately run the review step: trace at least one edge case by hand and confirm the code handles it. If it doesn't, write a precise refine prompt (symptom + cause + fix) rather than "fix it." The goal is to feel the full loop once, end to end, on code you can fully evaluate.
Prompt of the chapter
Act as my pair-programming partner. I'll describe a small, self-contained
function. Before you write any code, restate the requirements and list the
edge cases you'll handle so I can confirm them. Then write the function
plus unit tests covering those edge cases.
Function: <one-line description>
Inputs / outputs: <types and shapes>
Constraints: <stack, libraries to use or avoid, performance limits>
After the code, point out any assumption you made that I should verify.
Keep it to the simplest version that satisfies the requirements.