Case study, 2026
Cortex AI
Cortex AI is my private context-engineering system: it turns projects, notes, meetings, and decisions into inspectable context that agents can use without starting from zero.
Private system · public technical boundaries only
Context engineering
Cortex AI
Sources
Useful work
Problem
Generic assistants begin each task with a thin prompt and no durable understanding of the work around it. I wanted useful continuity without giving an agent unlimited authority.
Solution
Cortex compiles readable source material into linked context, gives each workflow a narrow operating brief, and places an explicit human gate in front of public or outbound action.
Impact
I use the system daily. This public explanation covers the model and boundaries without exposing personal data, providers, live endpoints, schedules, or private repository topology.
Build notes
- Python
- TypeScript
- React
- PostgreSQL
- Markdown
- whisper.cpp
- launchd
It is context engineering
Cortex is not a smarter chat window. It is the context layer behind the work I do with agents.
Projects, meetings, notes, people, decisions, and evidence remain readable files first. Cortex connects them, maintains an operating brief for each area, and gives an agent the smallest useful slice for the task at hand. The system matters more than any one model because the context can survive when the model changes.
The loop
- Sources arrive. A project changes, a meeting ends, a decision is made, or I capture a note.
- Context compiles. The change updates the several connected pages whose understanding actually changed, rather than disappearing into one folder.
- An agent does narrow work. It can draft a brief, inspect a system, propose an edit, or prepare the next step within explicit scope.
- A person decides. Public posts, messages, payments, and other consequential actions stop for human review.
That is the whole public diagram: sources → compiled context → useful work → human decision.
Why plain files still matter
The wiki is the source of understanding because I can read it without a special interface, inspect why a claim exists, and use normal version control to see what changed.
Structured databases support workflows that need state. A desktop interface makes the system easier to operate. Neither replaces the readable layer underneath.
Private surfaces stay private
The Content Studio belongs inside the private Cortex desktop app. It can organize ideas, recording days, edit batches, assets, and channel-specific drafts, but it is not part of this public website.
The same boundary applies across the system. This case study does not enumerate personal data, client information, live providers, endpoints, schedules, secret management, or internal failure notes.
Outbound means human
Cortex can help turn work into a draft and prepare an approved compose surface. It does not perform the final public action. I review the actual destination and click the final publish or send control myself.
That constraint is product design, not a temporary limitation. Context should make an agent more useful; it should not quietly make the agent the owner of my voice.
Current state
Cortex is in daily personal use and continues to change with the work it supports. Some workflows are local, some are scheduled, and some are intentionally manual. The honest measure is whether I can inspect the context, understand a failure, and remain in control—not whether the system looks autonomous in a demo.