Instructions

The Instructions screen is where an admin rewrites the standing instructions the assistant follows: its default persona, its rules for citing sources, how it formats answers, and how it plans and writes deep research reports. This page covers what you can edit, what you can't, and how changes take effect.

Where to find it

Open the user menu, choose System, then Instructions.

What you can customize

There's a fixed list of named instructions, grouped into three categories:

Auto

• Auto - Role Identity. The base instructions that define who the assistant is and how it behaves by default, before any specific agent's own instructions are layered on top.

Research

• Section Drafting. How individual sections of a research report get written: citation style, target length, and structure.
• Research Synthesis. How drafted sections get assembled into one finished document: transitions between sections and removing repeated points.

Shared

These apply across every mode, not just one:

• Grounding Rules. Citation format, the rule against stating things the assistant can't support from retrieved material, and how sources get attributed.
• Answer Layout. When the assistant reaches for a heading, a list, a table, a checklist, or a code block, versus when it should just write plain prose.
• Quality Check / Critic. The instructions used to grade a research report draft before it's finalized, across completeness, accuracy, clarity, actionability, and citation quality.

Each one opens in a text editor. Change the wording, save, and it applies immediately, everywhere that instruction is used. No restart is required.

TIP

Before you start editing an instruction you might want to walk back, copy the current text somewhere safe first. See "Default vs. Custom" below for why.

What isn't offered here

A handful of internal instructions aren't listed on this screen at all, and can't be edited through it. These are the ones whose output is read directly by the platform's own logic rather than shown to a person, so an unexpected rewording would break the mechanism that depends on it:

• The instructions that decide whether the assistant should search a knowledge base before answering at all.
• The instructions that name a new conversation.
• The instructions that read a newly uploaded document and extract its name, description, and tags.
• The instructions that pick which report shape and outline to use when a research job starts.

These stay fixed so the rest of the platform keeps working correctly. If you need different behavior from one of them, talk to your delivery engineer.

Default vs. Custom

Every instruction starts out as the platform's built-in wording, marked Default. The moment you save a change to one, it's marked Custom instead, and a small dot appears next to it in the list so you can spot at a glance which instructions have been changed from stock.

There's no restore to default action today. If you want to revert an instruction you've customized, you'll need to paste back the original wording yourself, which is why it's worth keeping a copy before you start editing something you might want to undo.

When to edit these

Common reasons to customize an instruction:

• Matching your organization's tone, terminology, or house style.
• Adding domain-specific guidance (for example, a compliance disclaimer that should appear in every research report).
• Tightening the grounding rules for a deployment with a low tolerance for unsupported claims.
• Adjusting report structure defaults if your teams consistently want a different shape than the built-in wording produces.

Because Grounding Rules and Answer Layout apply everywhere, a change there affects every conversation on the platform, not just one agent. Test after changes to shared instructions before rolling them out widely.

Next

• Report shapes for the separate catalog that controls how research reports get outlined.
• Roles & permissions for who can reach this screen.