Context is the new CMS
For twenty years teams ran a content management system. The next era runs on context management: what your AI knows before it writes a word.
For about twenty years, the center of gravity for software teams was the CMS. A content management system answered one question well: where does the content live, and who can change it. We built our whole craft around that question.
That question is quietly becoming the wrong one.
The new center of gravity is context management. Not where the content lives, but what the model knows about you before it produces a single word. The bottleneck moved up the stack, from storage to knowledge.
The one-line version
A CMS stores what you wrote. Context management stores what the AI needs to write as you.
Why this matters now
A raw model is a brilliant stranger. It writes fluently and confidently about your business while knowing nothing about it. The gap between a generic answer and a useful one is almost entirely context: your brand voice, your constraints, last quarter's decisions, the three things you never do.
You can close that gap two ways. You can paste the same background into every prompt forever, or you can build a system that carries it for you. The first feels productive. The second compounds.
What good context engineering looks like
The teams getting real leverage treat context as a first-class artifact, not an afterthought you stuff into a prompt. A few patterns I keep seeing:
- Knowledge lives once. Brand guidelines, product facts, and past decisions are added in one place and referenced everywhere, not re-explained per conversation.
- Voice is trained, not described. "Write in our voice" is a wish. A set of real examples the model can match against is a spec.
- Persistence beats recall. The system should remember across people and sessions, so the newest teammate gets the same context the founder has.
// Context is data you curate, not prose you retype.
type Context = {
brandVoice: Example[] // show, don't tell
facts: KnowledgeItem[] // the things that are simply true about you
guardrails: string[] // the three things you never do
}The trap
More context is not automatically better context. A model drowning in stale documents is worse than one given the five facts that matter. Curate ruthlessly.
The shift in one sentence
The question used to be "where is the content." The question now is "what does the model know before it starts." Teams that answer the second one well will make the first one feel like a footnote.
That is the bet we are making at Juma, and it is the lens I will keep coming back to in this blog.
Did this resonate?
Get new essays by email
Field notes on AI-native products, straight to your inbox. No spam, unsubscribe any time.