context engineering

Rod Johnson's July 2025 essay treats "context engineering" as a real advance over prompt engineering, then argues its definition is incomplete: it accounts for what is sent to the model, not what comes back, and ignores integration with existing business systems. His answer, Domain-Integrated Context Engineering (DICE), uses a domain model to structure both inputs and outputs, turning what he calls a delicate art into something that can be refined, reasoned about, and tested.

The Google whitepaper signed by Addy Osmani, Shubham Saboo and Sokratis Kartakis makes it the central skill: quality depends less on the prompt than on the context supplied. Six context types (instructions, knowledge, memory, examples, tools, guardrails) split between static context, always loaded and expensive, and dynamic context, fetched on demand. Agent Skills are the flagship dynamic pattern.

Hugo Lassiège shows what that looks like at the level of one developer's toolchain. Writing about production products whose code is now almost entirely generated, he frames his practice as «Context Engineering» rather than «vibe coding», which he attributes to «Andrej Karpathy» as experimentation. His context is stratified by loading moment: a short permanent `CLAUDE.md`, path-triggered rules, personas in `.agents/*.md`.

The term also gets positioned as a stage that is already passing. On BFM Business in May 2026, Rémi Jacquet dated a sequence: prompt engineering in 2024, context engineering in 2025, harness engineering in 2026. Lushbinary stacks it differently, as the middle layer under «Loop Engineering».