Capstone: fix a broken AI assistant
Diagnose the trace, then apply every fix you've learned until it's healthy.
1A shipped support assistant, failing. Tap each span to find what's wrong.your turn
Assistant health0% · broken
trace · support assistant · 1 request0/5 fixed
Tapped 0/5. Keep tapping spans, every one hides a defect.
→ continue← backR replay
You've reached the end
That's the whole arc, from a word, to attention, to agents, down to the GPUs. Now show someone what you can do.
Builds on7.2Observability: seeing inside5.2RAG: retrieval as a callback to similarity7.9Structured outputs: guaranteed JSON6.5Tools, permissions & trust7.3Evals: proving it works5.4Context rot: when more context hurts
Common questions
What is "Capstone: fix a broken AI assistant" about?
Diagnose the trace, then apply every fix you've learned until it's healthy.
What problem does it solve?
Here's a shipped assistant failing in five ways. Can you repair it?
What will I be able to do after this lesson?
You can diagnose a failing LLM feature from its trace and apply the right fixes: trim context, add retrieval, constrain output, set permissions, and gate on evals.