Context Engineering Is the New Prompt Engineering

In 2023 everyone wanted to be a prompt engineer. Screenshots of magic incantations. “You are a senior X with 30 years of experience, think step by step.” Five hundred characters of cleverness dropped into a chat window, and people acted like they’d discovered fire.

Three years in, that whole discipline looks quaint. Not wrong — just tiny. It was a hack against stateless models with nothing to work with. You compensated for missing context by stuffing the prompt.

The actual leverage was never in the prompt. It was always in the context pipeline behind it.

The Shift

A prompt is an API call. It’s the thin request. Everything interesting happens before the call: what gets retrieved, what gets remembered, what gets routed, what world model the model is pointed at when you hit enter.

Call that context engineering. Prompt engineering asked, “How do I phrase this so the model responds well?” Context engineering asks, “What does the model need to know about this entity, at this moment, for this decision — and how do I pipe it in reliably?”

One is wordsmithing. The other is systems design.

The Thesis Miessler Kept Stating Until It Clicked

Daniel Miessler has been saying the same thing for a decade in different shapes. His phrasing has sharpened over time: first SPQA (State, Policy, Questions, Actions) in 2023, then “The Future of Hacking is Context” in mid-2025, then Unified Entity Context as the name for the whole pattern.

The core claim: software verticals go away. What remains is the entity — a person, a company, a city — modeled as a single living context. AI sits on top of that context and acts. Every security tool, every HR tool, every dashboard you’ve used is a thin UI over a slice of an entity that should have been represented once, properly.

His dinner vignette makes it concrete. You’re eating, your agent pings: a drunk employee at your target company just posted that a new subdomain shipped with default creds. The agent asks permission to probe. You nod. It confirms. You authorize exploitation. Appetizers arrive. None of that story is about prompt cleverness. All of it is about a context pipeline that watches public signals, fuses them into a world model of the target, and surfaces the one thing that deserves your attention now.

Attack or defense, the shape is identical: whoever has the cleaner, fresher context wins. Anthropic’s 2025 threat report — one actor using Claude to hit seventeen organizations in weeks — wasn’t a story about a clever jailbreak. It was a story about an attacker who had built the context pipeline first.

If you’re a defender without one, your logs look like a wall. The gap between “state change in your environment” and “your system understands that change” is now the only metric that matters. Context speed. That’s the discipline.

What I’m Actually Building

I write about this because I’m building it. DAI — my Digital Assistant Infrastructure — is, in engineering terms, a context pipeline with an LLM glued to the top.

The model isn’t the product. The context machinery around it is.

Let me name the pieces, because this is where the abstract becomes real:

Entity model. At the center sits Telos — my explicit, versioned representation of me. Goals. Beliefs. Projects. Problems. Relationships. Budget. Health. This is the single source of truth about the entity. Everything downstream refers back to it. When something changes — a new goal, a revised project, a resolved problem — Telos changes first. Memory and retrieval propagate from there. Never the other way around.

Memory. Session continuity lives in memory/. It’s structured — references docs/, never contradicts it. The distinction matters. Long-term facts live in docs/. Short-term working state lives in memory/. Rendered views live downstream. If a plan changes, docs/ updates first, then memory, then anything generated from them. Top-down, always. That’s not a style choice — it’s what keeps the context pipeline non-contradictory when five subagents are reading it in parallel.

Retrieval and routing. The model doesn’t see everything. It sees what’s relevant, loaded when needed. A lightweight router reads the request, decides what context slice to load (DAI internals? the user’s life? a specific project?), and pulls only that. Drowning the model in a four-million-token soup is how you get worse answers, not better ones. Context is a scarce resource even when it’s technically cheap.

Lifecycle hooks. Thirty-nine hooks fire on events — before a tool runs, after a file is edited, when a session starts, when a PRD is written. Hooks are the connective tissue of the pipeline. They sync state, enforce invariants, notify the dashboard, and — crucially — never write back to the source of truth. They read the PRD, they don’t edit it. That asymmetry is load-bearing. One writer, many readers. The same rule that makes distributed systems work.

The algorithm. Every non-trivial task runs through a seven-phase loop: observe, think, plan, build, execute, verify, learn. The phases aren’t aesthetic. They’re the place where context gets assembled (observe), pressure-tested (think), and verified against reality (verify). The prompt at any given step is trivial. The surrounding phase is where the work happens.

None of this is prompt engineering. All of it is pipeline engineering.

Why Prompts Became Thin

Once entity context, memory, retrieval, and routing do real work, the prompt shrinks toward nothing. A phase instruction might be twenty words. The model is already pointed at the right slice of the world; it doesn’t need to be coaxed.

Compare two systems that ask the same question — “should I take this job?”

The 2023 version: a dense prompt with role-playing, constraints, chain-of-thought triggers, and three pasted documents. Maybe a good answer. Maybe not. No memory of the next conversation.

The 2026 version: a one-line question. Behind it, the assistant already knows the goals in Telos, the current projects, the monthly burn, the open offers, the partner’s career constraints, the last five conversations where I talked myself out of similar moves. The “prompt” is a thin API call. The context does the heavy lifting.

The second one isn’t smarter because of a better model. It’s smarter because the pipeline is doing engineering work the prompt never could.

What This Means If You’re Building

A few things follow directly.

Stop optimizing prompts in isolation. If you’re iterating on wording without iterating on what the model can see, you’re polishing the handle while ignoring the blade.

Treat context as a product. Version it. Validate it. Monitor it for drift. Sarit Tager at Palo Alto talks about this in the security frame — a unified data lake across code, cloud, runtime, identity. Same pattern. Coverage before risk. You cannot prioritize findings if half your repos aren’t in the lake. You cannot answer questions about yourself if half your life isn’t in your entity context.

Enforce one source of truth per fact. If two places claim to know the same thing, one of them will go stale, and stale context is worse than no context. Hierarchy is not bureaucracy. It’s what keeps a living system coherent when it grows.

Watch the freshness gap. Miessler is right — the gap between “state change” and “system understands it” is the real security metric. It’s also the real productivity metric. My assistant is only useful to the extent it knows what I did this morning.

And stop confusing “bigger model” with “better system.” A four-million-token context window is a storage unit, not a strategy. What you put in it, when, and why — that’s the engineering.

The Real Work

Prompt engineering was a stepping stone. It taught us to think about what we tell the model. Context engineering teaches us to think about what the model can know, structurally, before we say a word.

The next five years of real AI leverage won’t come from cleverer prompts. They’ll come from people who treat context as infrastructure — entity-first, hierarchical, versioned, always fresh. Pipelines over phrases.

That’s where I’m pointing DAI. That’s where everyone I respect is pointing their work. And it’s where I’d bet on if you’re deciding what skill to build next.

The prompt is the tip. The pipeline is the iceberg.