llmfn∘layers
From Prompt to Product
The interesting part of an LLM product is not the model. The model is the middle. What surrounds it — how you retrieve, structure, ground, remember, act, and measure — is what separates a demo from something that works in production. This site is a map of those eight layers.
The Progression
Product
Sets up the product, data, and philosophy before the first layer
Reveals: Natural language is invisible without any LLM layers
Search
Users find products by keywords
Reveals: Intent is ignored
Prompt
Understands the intent in natural language
Reveals: Output is inconsistent & unstructured
Schema
Shape of model output is consistent
Reveals: Answer is not grounded in real product data
Context
Responses are now grounded in facts
Reveals: Conversations are lost between turns
State
Remembers the full conversation history
Reveals: Preferences are lost when session ends
Memory
System builds knowledge about the user over time
Reveals: Cannot interact with the world outside the context window
Tools
The system can read from and write to the real world
Reveals: Failures are invisible and hard to detect
Evals
Output quality is measurable and testable
Reveals: Production ready. What you build next are product decisions
Reading Guide
Workshop participants
Coming to the workshop? Read layers 1–4 before the day. Each page takes about ten minutes.
Exploring on your own
Exploring on your own? Start anywhere. Each page stands alone and links to the next.