Pick your model like you’d pick a doubles partner.
I tested four LLMs head to head on the same micro bot task. Same prompt, same framework, same expectations.
The fast one (Grok) moved quick but left a code mess behind. GPT was painfully slow and so chatty I spent more time reading its commentary than reviewing code. Gemini started apologizing mid-session from all its errors. Claude was the only one that actually read my plan, followed it step by step, used the framework, and wrote the cleanest code of the bunch.
The lesson? Not all models are created equal for the task at hand. For bot making, Claude gets it.
Write the plan before you write the prompt.
This is the thing that changed everything.
I used ChatGPT to draft a full markdown plan with checkboxes. Then I handed that plan to the coding AI and told it to go section by section.
When I let the AI do the whole plan at once? Disaster. When I forced it to check in after each section? Cleaner code, fewer errors, actual progress.
Your prompt isn’t the strategy. The plan is.
Set rules like you’re onboarding a junior dev.
I created a rules file in my IDE that the AI had to follow. Five rules:
- Solve it simply. No over-engineering.
- Fix the problem. Don’t stack workarounds.
- Stay on task. Don’t wander.
- Suggest ideas. Don’t auto-implement.
- Report what you changed and why.
That last one is the sleeper. When the AI has to explain its changes, it catches its own mistakes before you do.
Point the AI at real docs, not its own memory.
I dropped links to python-sc2 docs, the framework docs, and open-source bot examples directly into the plan markdown file. Then told the AI to reference them before writing anything custom.
That single instruction cut hallucinated code in half.
LLMs are confident. Doesn’t mean they’re right. Give them source material and they’ll lean on it instead of making things up.
You don’t need to code. You need to know when the code is wrong.
Vibe coding doesn’t mean zero skill. You still need to understand your bot’s architecture well enough to spot when the AI is confidently heading in the wrong direction.
The AI won’t tell you it’s wrong. That’s your job.
I put all of this into practice building a full Terran 1-1-1 bot from scratch. It made it to ladder. Watch below to see the whole thing play out.
