vibe-coding-3-steps.md

How I do vibe coding in 3 steps (plus a Step 0)

0. Start with domain expertise
   Vibe coding works best when you have enough domain expertise to set direction, ask precise questions, and validate the output.

1. Jam with an AI chat — then “lock in” a single strong prompt
   I start by brainstorming with ChatGPT (or another chat model): goals, constraints, edge cases, and what “done” looks like.
   I end by asking the AI to produce one clean prompt that includes:

   • problem statement + scope (what to do / not do)
   • assumptions and constraints
   • acceptance criteria
   • suggested files/modules to touch
   • “verify it” requirements (tests, linting, manual checks)

2. Set guardrails in VS Code with custom instructions
   Before generating anything, I add one or more instruction files so the assistant follows my standards automatically (style, architecture, libraries, testing, security, etc.).
   I typically pull from the “awesome-copilot” instruction patterns and adapt them to my repo.

3. Use Plan Mode: paste the prompt, generate a plan, then execute
   In Copilot Chat Plan Mode, I paste the prompt from step 1 and ask VS Code to:

   • create a step-by-step plan
   • list files to change
   • call out risks/unknowns
   • propose checkpoints (build/tests) after each milestone and commit to git

   Then I iterate: adjust the plan, and only then let it implement.