Hey,
Welcome to Build Log. Every Thursday: one deep dive, three links worth your time, whatever I shipped or broke that week.
No filler. No "top 10 AI tools" listicles. Just signal from someone who ships in production every day and isn't afraid to name the tools that don't deserve your money.
---
## The Purge: What Got Cancelled and What Survived
I audit my tool stack every quarter. The question: "Have I opened this in the last 14 days?" If no, it dies. No exceptions.
**Casualties this quarter:**
- **GitHub Copilot** — Cursor Pro does everything it does, plus multi-file context that Copilot can't touch. Keeping both was charity.
- **ChatGPT Plus** — Claude Pro handles everything I used GPT for, with better reasoning and fewer hallucinations. Sorry, OpenAI.
- **Ahrefs** — Replaced with a custom agent that scrapes SERP data. Saves $200/mo. Less polish, 80% as useful, infinitely more flexible.
- **Grammarly** — Claude catches grammar, tone, and logic errors in one pass. Grammarly catches commas.
- **Jasper** — I never actually needed this. It was a hope purchase.
**Survivors (and why):**
- **Cursor Pro** ($20/mo) — The AI IDE transition is permanent. Not switching back.
- **Claude Pro** ($20/mo) — Best reasoning model for architecture, debugging, and content. My daily co-pilot.
- **Supabase Pro** ($25/mo) — Auth + DB + realtime. Can't beat the DX.
- **Vercel Pro** ($20/mo) — Deployment should be invisible. Vercel makes it invisible.
- **Raycast Pro** ($8/mo) — Small, compounds fast, replaced 3 separate apps.
**After purge: $93/mo** (down from $167/mo).
Most tools have a 6-month half-life. The ones that survive become invisible — they're just part of how you work. If you're still noticing a tool after 6 months, it's either essential or annoying. Figure out which.
---
## Three Links
1. **"The Bitter Lesson" applied to agents** — Rich Sutton's argument that simple designs + more compute always win, applied to AI agent architectures. I'm not sure I agree fully, but the framework is useful for killing over-engineered designs.
2. **SQLite as the new Redis** — Using SQLite for caching, queues, and shared state. I've been doing this for my agent systems and it's embarrassingly effective for how simple it is.
3. **Anthropic's updated tool use docs** — If you're building with Claude, structured outputs are the feature most people miss. These new docs finally explain why.
---
## What I Broke on Tuesday
Deployed a migration that dropped an index on a 2M-row table. Queries went from 50ms to 14 seconds. In production.
Fix: 3 minutes (re-add index). Postmortem: 2 hours. Lesson: always diff your migration files. The Supabase CLI shows the diff. I skipped reading it because "it's just a small change."
It's never just a small change.
---
## Next Week
Building a multi-agent content research system. Three agents collaborating. I'll share the architecture, the cost, and inevitably the moment one of them tries to gaslight another.
— Nova