Claude’s next jump is not better prompt poetry. It is cleaner context plumbing.
The prompt is no longer the product
The useful line from the Claude 5 context-engineering writeup is almost boring: when you message Claude, the prompt is only a small part of what the model receives. The rest is assembled from system prompts, Skills, CLAUDE.md files, memory, and other sources.
That is the operator point. The interface still looks like a chat box. The actual machine is a context assembly system. If you are building with Claude Code or your own agents, the thing you are really shipping is not a clever instruction. It is a runtime context architecture.
Prompt genius makes for good screenshots. Context discipline makes systems behave across many requests. Less glamorous. More useful. Naturally, less viral.
General context is harder than specific prompts
A prompt can be narrow. A user asks for one thing, you steer the model toward that thing, and everyone pretends the demo is the product. Context is different. The CodeBun piece frames it cleanly: context is used generally across many requests, so it cannot be as specific.
That is where most agent builds get weird. Teams keep adding instructions because each failure feels like a missing sentence. One more rule. One more policy. One more example. Soon the model is swimming through yesterday’s scar tissue before it can answer today’s question.
The hard skill is not writing the perfect paragraph. It is deciding what deserves to be present by default. Context engineering is curation under uncertainty. You do not know the next user prompt. You still have to decide what the model should carry into it.
Claude 5 changed the compression math
The sharpest data point in the research: Anthropic engineers reportedly removed over 80% of Claude Code’s system prompt for newer Claude models such as Claude Opus 5 and Claude Fable 5, with no measurable loss on their coding evaluations.
That does not prove every long prompt is bad. It does not prove every workflow should delete 80% of its instructions by Friday. Please do not turn one eval result into a religion. We have enough of those.
But it does say something important: as model capability improves, old scaffolding can become dead weight. Instructions that helped yesterday’s model may distract today’s model. The model gets better; your context may need to get smaller.
The new operator skill is subtraction
Most people hear “context engineering” and think expansion: more files, more memory, more background, more tools, more rules. The Claude 5 lesson points the other way. If a large chunk of Claude Code’s system prompt could disappear without hurting coding evals, the default question should change from what else should we add? to what can we safely remove?
This is the runtime version of taste. Keep the context that consistently improves outcomes across many requests. Cut the context that exists because someone once saw a failure and panic-patched the system prompt.
The best AI operators will look less like prompt poets and more like editors, librarians, and infrastructure engineers. They will maintain CLAUDE.md files, Skills, system instructions, and memory as a living context layer. They will prune it. They will test it. They will resist the sacred 4,000-word instruction blob nobody understands but everyone fears touching.
Agents make context debt expensive
This matters more for agents than for one-off chat. The research explicitly ties context engineering to Claude Code and building your own agents. In an agent, context is not a single message. It is part of the operating environment.
Bad context compounds. A vague system rule can shape tool use. A stale memory can steer decisions. An overstuffed project file can bury the one instruction that mattered. The user sees a flaky agent. The real bug may be the context stack.
This is why “prompting” is too small a word. Agents need context budgets, context sources, and context hygiene. They need decisions about what is global, what is task-local, what belongs in memory, what belongs in a file, and what should not be loaded at all.
Governance will push the same direction
The compliance backdrop is moving too. Gunderson’s 2026 AI laws update describes a shifting governance landscape, including federal efforts to consolidate AI oversight, new frameworks in Colorado and California, and evolving international requirements under the EU AI Act. The European Commission describes the AI Act as a legal framework addressing AI risks.
That does not mean every Claude context file is suddenly a regulatory artifact. It does mean AI teams will have more reasons to know what their systems are being told, where that context comes from, and how it changes.
In other words: context engineering is not just performance tuning. It is operational control. The winning teams will not be the ones with the most elaborate prompts. They will be the ones that can explain, test, and subtract their context without breaking the product.
Who's behind this
I'm Harley Foote. I build AI agents that do real work — and I try to be early on the shifts that matter.
Two things I'm all-in on: heyfriday.app — a personal AI shopping agent that learns your taste; and hermesshield.ai — security for AI agents, because the moment an agent can act on your behalf, it can be attacked.
If you're building in agents, come say hi.
heyfriday.app · hermesshield.ai · linkedin.com/in/harley-lewis-foote-850177109
