A Slidev presentation by Giorgio Galassi on the discipline of context engineering for AI coding agents — what agents know, when they know it, and how that knowledge survives session boundaries and tool switches.
LLM-powered coding agents are stateless by default. Every session starts cold. This talk makes the case that the highest-leverage skill when building with AI agents is not prompt engineering — it's context engineering: deliberately deciding what your agents know before the first prompt is ever written.
The talk covers:
- Why sessions fail (session amnesia, cross-project blindness, tool switch loss)
- The four named failure modes from the research community (context confusion, distraction, poisoning, clash)
- Empirical evidence: the U-shaped attention curve and context rot across 18 frontier models
- Four practical techniques: index-first loading, anchored iterative summarization, phase-based context loading, and just-in-time file retrieval
- The decision space: where memory lives, what storage primitive to use, and how to keep memory from becoming a liability
- Slidev — presentation framework
- Vue 3
- UnoCSS (Slidev default)
- Shiki — code highlighting
npm install
npm run devOpen http://localhost:3030 in your browser.
| Command | Description |
|---|---|
npm run dev |
Start local dev server with HMR |
npm run build |
Build static site for production |
npm run export |
Export presentation to PDF |
slides.md — all slide content and speaker notes
src/ — images used in slides
styles/ — custom CSS
All slide content lives in slides.md. Speaker notes are in <!-- --> comment blocks at the bottom of each slide.
Giorgio Galassi — Senior Frontend Engineer (Freelancer) · GDG Roma Città Organizer
ISC