Context Engineering Has Replaced Prompt Engineering as the Core Skill of AI Engineers
The Claim The craft of writing clever instructions has been superseded by the systems discipline of deciding what information an agent can access, in what form, and at what point in its reasoning — the practice now known as context engineering.
Why This Matters If prompt engineering is the core skill, the valuable investment is in writing better instructions. If context engineering is the core skill, the investment is in architecture: memory stores, skill files, tool selection, window management, and retrieval systems. These are different economic bets and different hiring profiles.
The Evidence
The shift is visible in how practitioners across the conference described their hardest problems. Brendan O'Leary at Kilo Code cited Andrej Karpathy's formulation of context engineering as 'the delicate art and science of filling the context window with just what needs to happen' and positioned it as the skill that separates teams getting genuine productivity gains from those stuck in the loop. His two identified failure modes — context bloat (quality degrades noticeably above ~50% window utilization) and context poisoning (stale, mixed, or exploratory content corrupting future outputs) — are both architectural problems, not instruction-quality problems.
Sally-Ann DeLucia at Arize gave the clearest statement: after nearly a year building Alyx, the team's primary learning was that 'context engineering matters more than prompt engineering for agent reliability.' The evidence she offered was operational: Alyx entered a vicious loop not because its instructions were wrong, but because the spans it generated filled its context, which caused failures that generated more spans. The fix was architectural — a head-plus-tail smart truncation strategy with a retrievable memory store, not better wording.
Pedro Rodrigues at Supabase provided the most controlled evidence. In a structured evaluation across four model/vendor combinations, he tested a baseline agent, MCP-only, and MCP-plus-skill on six real Supabase scenarios. The skill's specific contribution was a security constraint the model could reason about correctly only when it appeared in the context — and the improvement was consistent regardless of whether the underlying model was Claude Opus 4.6, Sonnet 4.6, GPT-5.4, or GPT-5.4 mini. The context content, not the model or the instruction style, was the lever.
Patrick Debois gave context engineering its own full lifecycle. The Context Development Lifecycle (CDL) treats prompts, rules, memory files, and skills as versioned, tested, packaged, and observable artifacts — with the same engineering rigour applied to context that software teams apply to code. His framing of 'context as fuel, LLMs as the engine' captures the inversion: the model is a commodity; the context is the competitive asset.
Nick Nisi's case study is perhaps the most striking. After generating 10,000 lines of skill content by chunking WorkOS documentation, eval times ran 68 minutes and results were poor. One skill caused a 20-percentage-point accuracy regression on a specific task. His solution: delete 95% of the content and replace it with 553 lines of focused 'gotcha' notes — precisely the landmines the model would otherwise fall into. Eval time dropped to 6 minutes and pass rates improved. The model already knew how to code; it only needed to know where the specific traps were in the WorkOS product.
The Nuance
The counterevidence is real but complementary. Matt Pocock's 'Grill Me' skill — which asks up to 100 adversarial questions before coding begins — is explicitly about improving instruction quality through dialogue. It reached 13,000 GitHub stars, suggesting genuine demand. Samuel Colvin showed GEPA prompt optimization lifting accuracy from 92% to 96.7% purely through wording evolution. These results do not contradict the context engineering thesis; they demonstrate that prompt quality matters within an already-well-constructed context. The relationship is hierarchical: context engineering creates the environment in which good prompts can work; without it, even excellent prompts are unreliable.
Verdict: Supported
Context engineering has superseded prompt engineering as the primary design discipline for production AI agents. Prompt engineering remains a meaningful skill, particularly for fine-grained optimization, but it is now a sub-discipline within the broader practice of context architecture.