Reading your own logs — why observability beats features

Most people add features. Better people watch the system.

The walkthrough

1. Most people add features. Better people watch the system.
2. If you can't see what your bots are doing, you can't improve them.
3. Three numbers. That's the whole dashboard.
4. Number 1. Latency. How long each call takes.
5. Plot it. Watch for spikes. A 30-second p99 means a bad model picked the wrong moment.
6. Number 2. Errors. How often calls fail.
7. Spikes are signals. Rate-limited? Bad prompt? Wrong model? Logs tell you.
8. Number 3. Cost per request. Average dollars per call.
9. Trending up = your prompts are getting longer or you're calling the expensive model too much.
10. Three numbers. One screen. Twenty minutes to build.
11. SQLite for logs. One HTML page. Plain JS fetch. No analytics SaaS needed.
12. Or wire it through n8n. Either way works. Don't pay $40/mo for Datadog yet.
13. Real example. Caught a 3am cost spike from an agent loop. Killed it before $200.
14. Red latency? Smaller model. Red errors? Better prompt. Red cost? Move to local.
15. This dashboard is what feeds the audit. Decisions go from gut to data.
16. Drop the dashboard.html in your project folder. Bookmark it. Open it daily.
17. Decisions get 10x better when you can see what's happening.
18. Stop adding features. Start watching the system.
19. localmodeldojo.com — Founder pricing $24/mo for life · code LAUNCH30 · 30 days only

— Jake Morris · Oklahoma veteran · localmodeldojo.com