The AI chat integration works out of the box, but two customizations let a team make the assistant work the way they work, without a code change. Both are set in Settings, on the “AI Assistant” tab, and both take effect immediately: a saved change is live on your next question, with no need to remove and re-add the connector.

Custom instructions: freeform guidance the assistant reads.
Saved prompts: reusable, named analyses you invoke instead of retyping.

Scopes

Both are scoped at three levels, and the most specific one wins:

Scope	Applies to	Who can set it
Personal	just you, in every project	any member
Project	everyone working in one game	Editor and above
Workspace	the whole team	Admin

When two scopes define the same thing, the narrower one takes precedence: your personal instruction overrides a project one, which overrides a workspace one. This is the same precedence as roles and permissions elsewhere in Asemic.

Custom instructions

Freeform text folded into the assistant’s context so it follows your team’s conventions. A few examples:

When I say "active users", use DAU.
Prefer chart-query over render-chart for ad hoc questions.
Revenue is net of refunds unless I ask otherwise.

Keep them short. There is a per-scope length limit (around 800 characters), and that is deliberate: short, specific guidance steers the model reliably, while long instructions dilute its attention. Project instructions are applied automatically whenever the assistant is working in that project, so you do not have to restate them.

What instructions do and do not do: they guide how the assistant works with your governed definitions. They never change the definitions or the methods themselves, which live in the data model, and they can never widen your access (see Safety below). Think of them as house style for the assistant, not new rules for the data.

Saved prompts

A saved prompt is a reusable analysis you invoke by name instead of typing it out each time. For example, a “Weekly monetization check” that always pulls the same metrics for the same window, or a “Launch retro” you run after each release.

A prompt can take arguments through {{placeholder}} slots, filled in when you run it:

Build a {{period}} retention retro for {{project}}: cohort curves,
the biggest week-over-week movers, and any annotations in the window.

Two tools deliver saved prompts, and both return fresh data on every call, so a prompt you just saved is usable right away:

list-saved-prompts: the prompts available to you (your personal, project, and workspace scopes), with their arguments.
run-saved-prompt: fills a prompt’s arguments and returns it for the assistant to execute against your governed metrics.

In practice you ask the assistant to run one by name (“run the weekly monetization check”), and it resolves the template and answers. Saved prompts follow the same three scopes and precedence as instructions.

Safety

Customization can never escalate access. Every tool still authorizes against your own role, so no instruction or prompt, including a workspace-level one set by someone else, can make the assistant return data you are not entitled to. The only residual risk is guidance that is misleading, not a data exposure.
Authoring follows your role: personal customization for any member, project scope for Editors and above, workspace scope for Admins. The same policy the write tools already use.
Shared instructions are team-authored guidance: they shape the assistant’s behavior, never its permissions.

Where to set them

Open Settings and select the “AI Assistant” tab. You will see an instruction editor for each scope you are allowed to change, and a place to create and manage saved prompts (argument names are picked up automatically from the {{placeholders}} in the prompt text).

Next: AI chat integration for the connect flow, or Roles and permissions for who can set what.