Strict JSON schemas cut function-calling errors more than prompt tweaks
Tightening a tool's JSON schema (enums, required fields, no free-form strings) reduced malformed function calls more than any prompt change did.
Welcome
The Product Bet
Long-form thinking on AI product management, voice AI, and RAG, plus short daily notes on whatever broke or clicked that day.
Featured Agentic Workflows
Strict JSON schemas cut function-calling errors more than prompt tweaks
Tightening a tool's JSON schema (enums, required fields, no free-form strings) reduced malformed function calls more than any prompt change did.
Explore by Topic
Jump straight to the kind of problem you're trying to solve.
Latest Articles
Long-form pieces on the systems behind AI products.
- Agentic Workflows
Designing Agentic Workflows: From Chatbot to Autonomous Assistant
A practical framework for deciding how much autonomy to hand an AI agent, and where humans still need to stay in the loop.
- RAG
RAG Pitfalls Every PM Should Know Before Shipping
Retrieval-augmented generation looks simple in a demo and falls apart in production. Here's what product managers consistently get wrong before launch.
Quick Reads
Short TIL notes, logged as they happen.
-
-
Cache embeddings by content hash, not document ID
Keying an embedding cache by document ID instead of content hash silently serves stale vectors after edits.
-
Streaming TTS hides latency if you flush audio early
Perceived latency in voice agents drops a lot when you start playback on the first audio chunk instead of waiting for the full synthesis.