Eight topics from the same beginner-friendly deck as our LLM track—how prompts steer models, classic reasoning templates (CoT, voting, tree search), structured ARQs, verbalized sampling for diversity, and JSON-first prompting for reliable automation.
Instructions—not weights—steer the model; good prompts elicit steps, constraints, and depth.
Code, math, logic, and multi-step tasks all benefit from the same reasoning patterns.
Step-by-step reasoning before the final answer—the most widely used nudge.
Many chains, one vote: pick the most common final answer.
Branch intermediate steps and search for the best path—not just the best final guess.
Structured queries (often JSON) replace improvisation; Parlant is one implementation story.
Recover pre-training diversity with a training-free prompt that asks for a distribution, not one answer.
JSON, XML, Markdown—what matters is structure; the PDF compares formats and closes with a summary.