I figured out the best way to vibe code

TL;DR

• 1

  Experts automate the whole coding loop: Matthew Berman says the real jump from beginner to expert is moving from manual prompting to systems that trigger agents automatically and keep running until a defined goal is met.

• 2

  Skills are the core reusable building block: His rule is simple, if you do something more than once, turn it into a skill, whether that is auto-review, test commands, GitHub issue formatting, or API-specific instructions.

• 3

  Greptile becomes an AI code reviewer with a confidence score: In his workflow, Greptile reviews every PR, summarizes changed files, scores merge confidence from 0 to 5, flags issues, and even generates prompts an agent can use to fix them.

• 4

  Loops make maintenance work continuous instead of manual: He shares concrete loops like an overnight docs sweep, a nightly production error sweep, and a performance loop that kept optimizing until every page, modal, and sidebar loaded under 50 ms.

• 5

  Cloud agents win when you run 10 to 30 agents in parallel: Berman likes local agents for speed and control, but says cloud agents are better for isolation, parallelism, and remote access, especially when his computer otherwise slows to a crawl.

• 6

  Merging and deploys are still a mess at agent scale: Even after talking with OpenAI, Cursor, and top agentic engineers, he says parallel PRs rebasing and re-triggering CI is still an unsolved bottleneck, with batch commits as only a partial workaround.