Vibe Coding Part 8: Every Mistake We Made (So You Don't Have To)

The Mistakes Nobody Shows You

Social media shows the highlight reel. "I built a SaaS in a weekend!" "AI generated my entire app!" What they don't show is the 3 AM server crash, the pricing bug that lost money for a week, or the GPU they bought that can't run the model they need.

Here's every mistake we made building SAVAH — and the systems that prevented us from making them twice.

Mistake 1: The Pricing Disaster

We uploaded products with estimated prices instead of real costs. For a week, we were selling items at a loss. The margin was negative on some products. We only discovered it when we manually checked the numbers.

The fix: A hard rule — every price must be calculated from real supplier cost + shipping + minimum 35% margin. No exceptions. No estimates. The formula is documented in our wiki and checked before every product goes live.

Mistake 2: The GPU Dead-End

We spent hours setting up Wan 2.2 for AI video generation. Installing dependencies. Downloading models. Configuring parameters. Then we discovered: our GPU has 4GB VRAM. The model needs 24GB.

The fix: Before committing to any AI model, check hardware requirements first. Not after you've invested hours. We now have a wiki/hardware/ section that documents every model's VRAM requirements.

Mistake 3: The Forgotten Configuration

PM2 was running. The site was live. Then we rebooted the server. PM2 didn't start. The site was down for hours before we noticed.

The fix: pm2 save + pm2 startup. Documented in the wiki. But we only documented it after the second time it happened.

Mistake 4: The Blog Image Violation

We used a logo as a blog featured image. It looked unprofessional. Blog posts should have real product images or editorial photos — not logos stretched to fill a banner.

The fix: Rule in AGENTS.md — "Blog featuredImage NUNCA debe ser un logo." Always use product images from the catalog or AI-generated editorial images.

Mistake 5: The Duplicate Knowledge

We re-discovered the same Cloudflare Tunnel fix three times. Each time, we thought "I'll remember this." We didn't.

The fix: The wiki system. Every factual discovery, every fix, every gotcha goes into the wiki immediately. Not "later." Not "when I have time." Now.

The Wiki System That Saved Us

Our wiki has 3 sections:

• Architecture decisions — Why we chose X over Y, with explicit rationale
• Gotchas and quirks — Things that aren't documented elsewhere (Cloudflare token recovery, USB path quoting, GPU limits)
• Operational knowledge — How to deploy, how to backup, how to recover from crashes

The wiki follows a strict format: situation → problem → solution → prevention. This means every entry is actionable. Not "Cloudflare Tunnel can break" (useless) but "If tunnel token is invalidated, rebuild from ~/.cloudflared/*.json and use --network host" (useful).

The 5 Rules That Would Have Saved Us Weeks

1. Document everything in the wiki immediately. Not later. Now. Your future self won't remember.
2. Check hardware requirements before committing to a model. 5 minutes of research saves hours of wasted work.
3. Never estimate prices. Always use real costs. A pricing error can lose real money.
4. Automate deployment early. Manual deploys are error-prone. Build the script on day one.
5. Set up PM2 auto-start before you need it. Don't wait for the first crash to configure it.

The Vibe Coding Mindset, Revisited

Vibe coding isn't about being perfect. It's about being systematic. Every mistake becomes a rule. Every rule goes into the wiki. Every wiki entry prevents the next mistake.

The AI doesn't make these systems. You do. The AI writes the code. You write the rules that keep the code correct.

What We'd Tell Our Day-One Self

• Start the wiki from hour one. Not week three.
• Test with real data early. Fake prices, fake products, fake reviews — they hide real problems.
• Set up monitoring. Disk space, process health, error logs. You can't fix what you can't see.
• Accept that AI makes mistakes. Your job is to catch them. That's the "vibe" in vibe coding.
• Ship early, fix fast. A live site with bugs teaches you more than a perfect local project.

SAVAH isn't perfect. But it's real, it's live, and it's getting better every day. That's the vibe coding journey: messy, iterative, and worth it.