Advanced Prompt Engineering

Before you can engineer prompts expertly, you need to understand what is actually happening when a model responds. This lesson explains the token-level mechanics that make every technique in this course work — and why some prompts succeed where others fail.

Chain-of-thought is the most studied prompting technique in the research literature. This lesson goes beyond the basics: how it actually works at the token level, the research statistics behind it, when it degrades performance, and how its effectiveness varies significantly across models.

When one answer is not reliable enough, generate several and vote. Self-consistency replaces greedy decoding with a sampling strategy that finds the most robust answer across multiple reasoning paths — and the cost is almost always worth it for high-stakes tasks.

Tree-of-Thought generalises chain-of-thought from a single path to an explicit search across multiple branching reasoning steps. For problems where backtracking matters — planning, puzzle-solving, multi-constraint design — it dramatically outperforms linear CoT.

A new category of model — reasoning models — has changed what prompting means. Claude Extended Thinking, OpenAI o1/o3, and Gemini Deep Think all perform internal multi-step reasoning before producing output. This lesson explains how they work, how to prompt them correctly, and critically, when not to use them.

Before solving a specific problem, step back and identify the general principle it falls under. This two-phase technique from Google DeepMind produces accuracy gains of 7–36% across STEM, temporal reasoning, and knowledge retrieval — by grounding answers in first principles before engaging with details.

Few-shot prompting in Track 2 introduced the concept. This lesson goes deeper: how many examples is optimal, how to select them, how ordering affects accuracy by more than 40%, the many-shot regime enabled by large context windows, and how to construct contrastive example pairs.

A production system prompt is load-bearing code, not a polite request. This lesson covers the professional structure that controls how models behave, the instruction hierarchy that governs what overrides what, model-specific differences, and why negative instructions reliably fail.

A well-designed persona shapes everything from vocabulary to reasoning style. This lesson covers the five dimensions of persona consistency, how contextual priming shapes model behaviour before the first user message, persona drift prevention in long conversations, and why pattern priming is a more precise tool than role assignment.

Complex tasks that exceed what a single prompt can reliably produce are solved by chaining: breaking the task into sequential steps where each output feeds the next. Research shows +15.6% accuracy over monolithic prompts. This lesson covers the four chaining patterns, validation gates, and failure handling.

Meta-prompting uses AI to write and improve prompts for AI. When formalised with a framework like DSPy, it becomes Automatic Prompt Optimization — a systematic, data-driven approach to prompt development that outperforms manual crafting on measurable tasks. This lesson covers both the technique and the tooling.

Models can critique and improve their own outputs — without external feedback or additional training. Self-Refine achieves ~20% absolute improvement across code, writing, and reasoning. This lesson covers the three-role loop, Chain-of-Verification, constitutional self-correction, and where these techniques reach their limits.

Getting models to reliably produce parseable, schema-valid output is a solved problem — if you know the right method for each provider. This lesson covers failure rates by provider, the differences between Claude tool use, OpenAI structured outputs, and Gemini responseSchema, plus constrained decoding for self-hosted models.

Prompts that are not tested are not trusted. LLM-as-judge achieves 80% agreement with human evaluation at 500x lower cost. This lesson covers how to build eval suites, the known biases that corrupt automated evaluation, regression testing in CI/CD, and the A/B testing patterns that replace intuition with data.

Prompts in production need versioning, cost controls, and security hardening. This lesson covers semantic versioning for prompts, the optimisation stack that achieves 60–80% cost reduction, prompt debugging by bisection, and the security techniques that reduce jailbreak success from 87% to under 4%.

Prompt engineering is maturing from craft to discipline. This final lesson covers agentic prompting patterns, memory management for long-running agents, building a team prompt practice, and where the field is heading — including the honest case for why prompt engineering remains essential in a world of reasoning models.