Engineering

Effective Prompt Engineering

Prompt engineering without the folklore. A practitioner's tour of what the evidence actually says about prompts — as specifications, as speech acts, as artifacts that fail under complexity and decay as models improve. Covers chain-of-thought, few-shot, persona, structured output, context engineering, systematic optimization (DSPy, MIPRO), empirical methodology, and the engineering judgment that holds it all together.

10 modules

Prompts Are First-Class Artifacts

Shifting from chat messages to production specifications — and why it changes everything about how you write them

How LLMs Process Language

Why your obviously clear prompt produces garbage — a linguistic model

Few-Shot Prompting: Evidence and Limits

What the empirical record says about example count, over-prompting, and when your prompt stops generalizing

Chain-of-Thought: Evidence and Limits

What the research actually says about when CoT works, when it fails, and why its output cannot be trusted as a debugging signal

Structured Output and Schema-Driven Prompting

Make your prompts behave like typed interfaces — testable, verifiable, and maintainable

Context Engineering and Chunking

The information you give a model is as engineered as the instruction you write

Persona Prompting, Constitutional Constraints, and Injection Risks

The 'act as an expert' pattern is everywhere. Here is what the evidence says about when it works, when it silently hurts you, and how to build prompts that hold under adversarial pressure.

Empirical Methodology for Prompts

From prompt folklore to engineering discipline: how to test what actually works

Systematic Prompt Optimization

From manual iteration to automated pipelines — and knowing when not to cross that line

Portability, Limits, and Engineering Judgment

How to make durable decisions in an unstable discipline

How this plan was made

Each plan on learnings is built by a hand-crafted agentic pipeline: research agents gather primary sources, a claim reviewer verifies facts against them, and a sequencer orders modules for how people actually learn. The curation — topic selection, framing, editorial standards — is Nicolas's. The research and writing is AI-assembled.