Context engineering

Context engineering for a company, not a prompt

Context engineering is deciding what an AI sees before it acts. For one agent that is its window; for a company it is one connected truth that every agent and every colleague reads, kept right by the people who know.

Definition

What context engineering is

Context engineering

Deciding everything an AI model sees before it acts: its instructions, the knowledge it is handed, the tools it may use and what it remembers. Prompt engineering words one instruction; context engineering decides the whole window.

The term grew out of building agents. An agent on a long task needs the right rules, the right facts and the right tools in its window, and little else: material that does not belong there can make the answer worse.

For one agent, that is a design problem. For a company, it is a question of where the facts live and who keeps them right.

Why more context is not the answer

Piling in pages gives you versions, not the truth.

The first answer to a model that does not know enough is to give it more: paste the brand guide, attach the strategy deck, point it at the wiki. The window fills with documents that overlap and contradict each other, and the model has to guess which one holds.

What an AI makes depends on what it knows, so the work moved from making the thing to keeping right what it is made from. One living truth, not more context.

Cards, notes, a folder and people pile up in a turning cloud. As it turns, each piece travels to its place in the brain's map and becomes a sphere of its group, every relation a thread, and the map draws together into the Epic Context mark.
  1. Everything you know, in pieces.
  2. Everything in its place, connected.
  3. One brain.

An illustrationNot more pages: every piece in its place, linked to what it rests on. Made from our own brain, in the suite's Content app.

Context for a company

Many people, many tools, one truth.

In a company the context is not one window. A product owner briefs Claude, a developer briefs Claude Code, an agency briefs its own tools, and every one of them works from a copy. The copies drift.

The answer is relations: every fact linked to what it rests on, so an agent can follow a feature to the job it serves, the decision it came from and the KPI that measures it.

Every fact linked: one feature and what it rests on
A featureLinked to
serves the job ofUsersthe persona it is for
came fromBuildthe decision and its owner
speaks inBrandthe tone of voice
lives onProducta page in the information architecture
is measured byPerformancea KPI that ladders up to the OGSM

How Epic Context does it

A briefing per task: the right expert and exactly the facts it needs.

Before an agent writes, it asks for a briefing. The briefing names the expert for the task and hands over the blocks that kind of task must see, in full, then related blocks as excerpts up to a token budget, each with its version.

  • A briefing per task. A screen, a campaign, a strategy question or a backend change each get their own.
  • The right expert. Eight experts, one per group, each with sub-skills, and your company's own rules on top of the method.
  • Read first, then write with a reason. A write gate checks that the agent read before it wrote: off, warn or enforce.
  • Every write lands in the changes feed with its source and reason, and can be undone in one click.

Epic Context · Architecture

One brain, layer by layer

  1. Your workspace

    Where you already work

    Colleagues and agents work together in the Epic Suite, in your own tools, or in an AI workspace you connect. You ask the question and decide what happens with the result.

    • Epic Suite
    • Claude
    • Claude Code
    • Cursor
  2. Agents and models

    The thinking

    Agents do the work when you ask. Every model call goes through one door that counts it. Today that model is Claude; the door can point an organisation at another model, and a second model has not run yet.

    • One door
    • Claude today
  3. One gateway

    One way in

    The AI and the brain talk through one connector, with your own company account: MCP for Claude, Claude Code and Cursor; an API with keys, bundles, episodes and webhooks for apps; a plugin for Claude Code.

    • MCP
    • API
    • Plugin
    • OAuth
  4. This page is about this layer: Context engine

    The briefing

    Every task gets the right expert and exactly the blocks it must see, in full, then related blocks as excerpts up to a token budget. Read first, then write with a reason.

    • A briefing per task
    • Write gate
  5. Knowledge

    Craft and memory

    Craft is the experts: one per group, each with sub-skills, and your own rules on top. Memory is the brain: eight groups, thirty-four kinds of knowledge, every fact linked to what it rests on. The editors order it every night and propose; they never rewrite.

    • 8 experts
    • 8 groups
    • 34 kinds of knowledge
  6. Foundation

    Yours

    Every change has a reason, a source and an owner, and can be undone. Locks refuse every writer. You can export the whole brain at any time; running it on your own server is planned.

    • Export any time
    • Own server: planned
The briefing layer sits between the gateway and the knowledge: it decides what each task sees.

Where it stops

The experts bring the method. Your company brings the facts.

Context engineering decides what an agent sees. It does not decide what is true: people do.

Where context engineering in Epic Context stops
What stands in the wayWhat we bringWhere it stops
Every AI works its own wayExperts: the right way of working for each task, growing with what the company learnsThey bring the method; the company brings the facts
Nobody dares to let AI write to what is trueEvery change has a reason, an owner and an undo; locks for what must not moveA person still decides what holds
Knowledge goes stale when nobody tends itEditors that connect, signal and propose, and never rewriteSomeone has to answer their proposals

Questions

What is context engineering?

Context engineering is deciding what an AI model sees before it acts: its instructions, the knowledge it is handed, the tools it may use and what it remembers. For a company it also means deciding where that knowledge lives and who keeps it right, so every agent reads the same truth.

How is context engineering different from prompt engineering?

Prompt engineering words one instruction. Context engineering decides everything else in the model's window: which facts, which rules, which tools and which history. A good prompt on the wrong context still gives the wrong answer.

What is context rot, and how do you avoid it?

Context rot is what happens when a model's window fills with too much, too old or contradicting material, and the answers get worse. You avoid it by giving a task the blocks it needs, not every page: a briefing hands over the required blocks in full and the rest as short excerpts, up to a budget.

Do I still need a CLAUDE.md or AGENTS.md?

Yes, for the repo's own rules: how to build, test and name things. The company's brand, users, design system and decisions do not belong in a repo file; they come from the brain through one connector, the same for every repo and every tool.

Which models does this work with?

Any AI client that speaks MCP can connect. Claude, Claude Code and Cursor are the ones we test and use every day.

Start with one real project

Epic Stories installs Epic Context with your team, on your own work, and then steps back. Tell us where you would start.