Context engineering is the job now
Prompt wording is a rounding error next to what the model can see and when. The discipline that replaced prompt engineering, and how we practice it.
Somewhere in the last two years, the craft quietly moved. The teams shipping reliable LLM systems aren’t winning on prompt wording — they’re winning on what the model can see at the moment it acts. The instructions, yes, but also: which retrieved passages, which tool results, which slice of history, which schemas, in what order, at what budget. That’s context engineering, and it has eaten prompt engineering as thoroughly as SQL tuning ate “writing queries.”
The tell is where debugging time actually goes. When an agent misbehaves, the fix is almost never a better adjective in the system prompt. It’s discovering the context window contains three conflicting policy versions, or a tool result shoved the critical instruction past the model’s useful attention, or history summarization dropped the one constraint that mattered. The prompt was fine. The context was a landfill.
Curation beats accumulation
The instinct with big context windows is to fill them — more documents, more history, more tools, more safety. Our evals say otherwise, consistently: six sharply relevant chunks beat fifteen loosely relevant ones; a scoped tool set beats the full catalog; summarized history beats verbatim replay past a surprisingly small window. Every irrelevant token isn’t neutral filler. It’s noise competing for attention with the signal you paid to retrieve.
The window is a budget. Spend it like one: allocate before you assemble — so much for instructions, so much for evidence, so much for history — and force each component to earn its allocation through re-ranking, compression, or exclusion. “Everything might be useful” is how you end up with contexts that are expensive and worse.
Context is a build artifact
The mental shift that changed our practice: stop treating context as whatever accumulates by the time the model is called, and start treating it as something assembled deliberately, per task, by a pipeline you own — with the same engineering hygiene as any build.
Concretely, that means the assembly is versioned and diffable: retrieval settings, ordering policy, compression rules, tool-schema selection — all config, all reviewable, all eval-gated in CI like any other change. When behavior shifts, we diff contexts, not just prompts. In promptev this became a first-class product object — Context Packs, reusable bundles of sources, rules, and access scopes — because watching users hand-assemble the same context per agent, badly, made the abstraction obvious. RBAC lives on the pack: identical question, different requester, different context assembled. Structure, order, provenance, permissions — engineered, not accumulated.
The skill that transfers
Prompt phrasing tricks rot with every model generation. Context discipline compounds: relevance, freshness, dedup, ordering, provenance, least-privilege scoping. Models keep getting better at using what they’re given — which raises, not lowers, the return on giving them the right things.
I still write prompts, of course. Ten percent of the job, maybe. The other ninety is deciding what the model deserves to know. Job title hasn’t caught up. The work already has.