05 section

Prompting and Context

Getting reliable behavior out of a model you cannot retrain — instruction design, example selection, reasoning elicitation, context budgeting, schema-constrained output, automatic optimization, injection defense

8 pages 26 min 5,256 words

01 the section

What is in here

The first four pages build up reasoning technique, from plain instructions to branching search. Context engineering and structured generation are where prompting stops being craft and becomes system design; DSPy answers the question of who maintains all these prompts once there are two hundred. Read the injection page if any untrusted text ever reaches your model.

01 3 min Prompt Engineering Fundamentals Prompts as code: instruction hierarchy, delimiters, role framing, and the intent-versus-constraint balance that decides how much a model's output varies run to run promptingfundamentals 02 3 min Few-Shot and In-Context Learning How many examples to show, how to pick them dynamically per query, and why the hard edge cases teach more than the clean ones promptingcontextretrieval 03 3 min Chain-of-Thought Reasoning traces moved from a prompt trick to a model feature; when to elicit them, how to verify them, and when overthinking hurts accuracy promptingreasoning 04 3 min Tree-of-Thought Branching search over reasoning paths — propose, evaluate, backtrack — and where MCTS-style exploration earns its large latency and token bill over a linear chain reasoningplanningprompting 05 5 min Context Engineering With million-token windows the question is no longer what fits but what earns its place: budgeting, lost-in-the-middle, thinking budgets, and cache-friendly prompt layout contextcachingcost 06 3 min Structured Generation Getting machine-readable output every time — JSON schema modes, constrained decoding with grammars and regex, multi-stage extraction, and what to do when validation still fails structured-outputpromptingtool-use 07 3 min Prompt Optimization with DSPy DSPy treats prompts as weights an optimizer tunes: signatures, modules, teleprompters, and the metric you must define before any of it works promptingevaluationproduction 08 3 min Prompt Injection Defense Direct and indirect injection, why input sanitizing alone never holds, and the architectural defenses: dual-LLM isolation, tagged untrusted spans, and output filtering securitypromptingreliability