Originally posted on LinkedIn
Vibe-coding, but done properly this time — Claude Code used right 🔥
After the hackathon I learned first-hand what to do and not do when using Claude Code.
For Ignia we’re building tech that takes our programs’ experience (and those of our partners like Sistema B… and another one we’re announcing this week 😏) to the next level.
We’re building Ignacio 🔥, an agent that can ‘mentor’ program participants in real time.
They can interact with him via web or WhatsApp, he has context on their projects, and he can give expert advice based on whatever problems they’re facing.
With Claude Code, in ~4 sessions of ~3 hours each, I’ve done what would have taken me a month (me plus a junior dev!).
Tips: 1️⃣ Have a .claude folder with context files. I keep CLAUDE.md, SPECS.md, and PLAN.md — to give it context on the project, technical specs, and the work plan. 2️⃣ First thing before anything else, write a very detailed CLAUDE.md (and SPECS.md if you already know what tech you want to use), and run “ultrathink an implementation plan” to put into PLAN.md. Review and edit the output thoughtfully. 3️⃣ Only build one thing at a time. Once it’s done, /clear and move on to the next. 4️⃣ Use Plan Mode a lot. When you’re happy with the plan, let it run.
Here’s master Ignacio. Built with Nano Banana and animated with Flow/Veo3.
Anyone have more tips? Is anyone curious about how to use Claude Code? Let me know in the comments or DM me and we’ll see if we can set up a session or workshop.
PS: I decided to use OpenAI’s Agents SDK and it’s incredible. Too bad I hadn’t seen it before the hackathon.