1 Mistake I Made When I First Started Vibe Coding And 3 Tips To Fix It

When I first started vibe coding, I made this huge mistake: I let the AI do the thinking for me.

I would give it a one-liner and ask it to build it. In hindsight, this wasn’t going to work. The AI won’t make the same decisions you’d make. It goes off the rails more often than not.

Tip #1: Come up with a rough plan

I take a few minutes to figure out what I want and how I’d build it without vibe coding. I give specific instructions on what, where, and how to do it. I tell the AI where to look in my codebase for reference or changes.

Tip #2: Ask the AI to plan first

I instruct it not to build yet. I say this — “Do not write any code yet. Plan first and wait for my confirmation.” I ask it to save the plan to a markdown file. For tricky bugs or features, I ask the Oracle, an Amp subagent that’s slow but good at thinking.

If I like the plan, I instruct it to build. If not, I iterate. But even with an agreeable plan, I don’t always get a solution I like. So…

Tip #3: Don’t be afraid to start from scratch

Sometimes, the AI throws slop at me. After a few follow-up prompts, it fails to zone in on a solution. When that happens, I stop and throw away everything it did. Yes, I lose tokens, but I’d rather have a working solution.

I then try to come up with a more concrete, less ambiguous prompt. Sometimes I get an acceptable solution. Sometimes, after two or three retries, I give up vibe coding and do it myself.

It took me a lot of hours, prompting, and tokens to figure these out.