The chief-of-staff thread: one persistent place to hold context
Don't start with a giant prompt library.
Start with one persistent chief-of-staff thread.
The job of that thread isn't to answer questions. Its job is to hold context, then route that context into action when you need it.
Most knowledge work fails in the seams: the Slack thread nobody copied into the brief, the customer objection that never reached the roadmap doc, the meeting note that contradicts the plan, the decision that got made but never written down anywhere, the draft that looks polished but ignores the latest constraint someone mentioned three days ago.
A good chief-of-staff thread becomes the place where those seams get collected, reconciled, and turned into next actions you can actually assign.
The pattern is already emerging inside Codex usage. OpenAI's own PMM example described using Codex for non-coding work: tracking context, understanding what's changing, turning scattered information into docs, and keeping cross-functional work aligned across teams.
That's not AI writing.
That's context operations, and almost nobody gets it right on the first try.
Here's how to structure the chief-of-staff thread.
One: give it a standing brief
Not a cute system prompt. A practical operating brief that covers what projects matter right now, what decisions are open, what sources are authoritative, what not to touch without approval, what format handoffs should use, and what "done" means for you or your team.
The thread should know the difference between a draft, a recommendation, a decision, and a task. If it can't tell those apart, you'll spend more time correcting outputs than getting value from them.
Two: make it an intake router
Every day, feed it the raw context that would otherwise decay: meeting notes, customer snippets, product changes, unresolved Slack decisions, email commitments, document updates, metrics or pipeline changes.
The first output shouldn't be prose. It should be triage: decision needed, action needed, waiting on someone else, context only, possible future idea, discard.
That single taxonomy prevents the most common AI failure, which is turning every input into another document you didn't ask for.
Three: let it spawn project threads
The chief-of-staff thread shouldn't do everything itself. It holds the map and spins out focused work.
"Create a project thread to draft the launch narrative."
"Create a background thread to compare the three customer segments."
"Continue the existing pricing thread and update it with this new objection."
OpenAI's Codex docs explicitly describe asking Codex to manage threads: find a related thread, continue an existing thread, pin or archive threads where available.
That's the key behavior. The chief-of-staff thread is the memory and routing layer. Project threads are execution lanes that report back to it.
Four: require handoffs back to the center
Every project thread should return the same shape: decision, why, evidence, open questions, recommended next action, artifact path or link.
Without that discipline, parallel work turns into parallel confusion, and you lose track of what got decided in which thread.
Five: pin the living context
Some threads are temporary. Some are infrastructure you return to every week.
Pin the ones that repeatedly absorb context: company strategy, customer voice, launch planning, board and investor narrative, hiring pipeline, weekly operating review.
Archive the rest aggressively before they become noise.
This is the discipline most people skip. They want AI to remember everything, but they never give it a clean place to remember from, then they wonder why the outputs drift.
The future of AI for knowledge workers won't be won by whoever has the longest prompt.
It'll be won by whoever has the cleanest context spine.
Start with one persistent chief-of-staff thread.
Then make every other thread report back to it.
---
Codex for knowledge workers is a five-part series:
- Codex is becoming a workspace, not a chatbot
- The chief-of-staff thread (you're reading this)
- Worktrees for knowledge work
- Automations turn follow-up into managed background work
- The Codex operating manual for knowledge workers
Setting up the tooling: the Codex CLI config.toml Guide 2026 covers profiles, permission modes, AGENTS.md, and plugins. The earlier Codex CLI Developer Guide covers the 0.42.0 command surface these workflows were first built on.