Track 3 · Prompts · Amber
Prompt & context engineering
Production prompt design is not clever wording—it is instruction hierarchy, versioned system templates, and deliberate context assembly under a token budget. This track teaches how to write durable contracts (role, rules, output shape), slot in RAG chunks and tool results from Track 2, defend against injection, and ship prompts with eval gates. Assumes you can call an LLM API and have built or read a basic RAG loop.
Guides in this track
Six deep-dive chapters. All guides are live—read in order.
Reading order: Prompt explained → Core prompting techniques → System prompts & instruction design → Reliable output formatting → Context window management → Advanced prompting in production
-
01
Prompt & context explained
Why prompting matters, mental model, prompt anatomy, instruction hierarchy, sensitivity, context in products, production checklist.
-
02
Core prompting techniques
Zero-shot, few-shot, chain-of-thought, role prompts, delimiters, negative instructions, and when each technique earns its tokens.
-
03
System prompts & instruction design
Versioned policy files, refusal rules, developer role, escalation paths, and testing prompts before deploy.
-
04
Reliable output formatting
JSON mode, schema constraints, citation formats, structured extraction, and parsing failures in production.
-
05
Context window management
Token budgets, lost-in-the-middle, summarization, history trimming, and ordering evidence for maximum recall.
-
06
Advanced prompting in production
Prompt versioning, A/B flags, regression evals, injection defenses, multi-model routing, and observability.