Personal Assistant

Bundle source: docs/ai-experts/personal-assistant.yaml

Unlike the persona docs on the other pages here, the Personal Assistant is a deployed Cosmos expert, not a prompt you paste. It appears in the Cosmos launcher for everyone in the workspace, and its definition — including its full system prompt — is version-controlled in this repo.

What it’s for

Sessions end, context doesn’t. Without persistent state, picking work back up means asking a new session to read an old one — a pattern that accounted for roughly a fifth of our Cosmos sessions, some of them thousands of messages long, purely to reconstruct where things stood.

The Personal Assistant keeps that state outside any single session:

• Task tracker — tasks.md, with Active / Done / Log sections and IDs T-001, T-002, …
• Memory — durable preferences and conventions it learns as you work
• Agent dispatch — it can launch other deployed experts as workers on your behalf and record the launch against the task

Open a new session and it reads your task list first, so it starts already knowing what you were doing.

Shared definition, private state

One expert, one prompt, everyone in the workspace — but your task list is yours. State lives in your personal Cosmos filesystem, so no one sees anyone else’s tasks or memory. This isn’t configured; it’s how the user-scoped filesystem works.

State also follows you across execution targets. A session on your local daemon and one on a cloud VM read and write the same task list.

Using it

Pick Personal Assistant in the Cosmos launcher and talk to it normally:

"Track this: PR 203 action failures, blocked on the missing keywords file."
"What am I working on?"
"Kick off a branch and implementation for the JSON schema roadmap item."

It has no triggers — nothing fires automatically. It runs when you open it.

Configuration

Field	Value
Visibility	Workspace-wide
Model	GPT-5.6 Terra
Reasoning effort	Platform default — not set in the bundle
Capabilities	Web access, GitHub
Environment	Default (includes this repo, so .augment/ rules and skills load)
Workers	Any deployed expert, including Issue Management Expert
Triggers	None

The bundle does not set reasoning effort because the current CI bundle schema does not recognize that field. GPT-5.6 Terra uses the platform default.

The system prompt is ours

The prompt is written out in full in the bundle rather than pulled from the platform’s stock template. It’s adapted from that template, with our own repository conventions added and the template’s Slack section removed, since no messaging integration is connected here.

This means the prompt is reviewable in a PR diff like any other code, and it doesn’t change unless we change it. Two small pieces are still pulled from the platform at session start — the shared memory-loading and memory-capture behavior — so those stay current without us maintaining them.

When a messaging integration is connected, add the notification instructions and the matching capability back to the bundle in the same PR.

Changing it

Edit docs/ai-experts/personal-assistant.yaml and open a PR. The Sync AI experts workflow validates and dry-runs the bundle on the PR, then applies it on merge to main.

Two fields are effectively permanent:

• slug is immutable after the first apply.
• Removing the file does not delete the deployed expert — do that deliberately with auggie cloud expert delete.

See also: AI Experts for the persona library, and Onboarding → Using Augment Cosmos for day-to-day workflow.