# Agent instructions

Standing decisions the agent follows in every conversation, set for you, a plan, a project or the whole company.



**Instructions** are standing decisions the agent should follow. They are the place for the
correction you keep making:

```text
Mobile ships after web. Stop proposing mobile work in the first pass.
We do not use feature flags for anything customer-visible.
```

Unlike [Context](/docs/context-sources), which is material the agent *reads*, instructions are
rules it *follows*. They are injected into every turn, so they shape the decision before the agent
starts thinking rather than sitting in a document it might consult.

## Four levels [#four-levels]

| Level         | Applies to                                  |
| ------------- | ------------------------------------------- |
| **You**       | Only your conversations, in this workspace. |
| **This plan** | Only that plan.                             |
| **Project**   | Every plan in that project.                 |
| **Company**   | Every plan in the workspace.                |

They are **additive**, not overriding. The agent sees all four, narrowest last. A genuine conflict
between two levels is yours to resolve, not something the system silently picks a winner for.

Your personal instructions are per workspace, so if you plan for two companies, each gets its own.

## Keep them short [#keep-them-short]

Instructions cost space in every single turn, so there is a cap on how much is injected. Write
decisions, not documentation:

* Good: "Always include a rollback step for anything touching billing."
* Bad: three paragraphs explaining your billing architecture. That is [Context](/docs/context-sources).

## The agent writes them too [#the-agent-writes-them-too]

As you plan, the agent notices how you work and writes instructions at the You, plan and project
levels on its own. That is why the project instructions grow over time without anyone opening the
settings screen.

## Company instructions need approval [#company-instructions-need-approval]

Company-level instructions reach every member's conversations, so they are **owner-only** and the
agent can only propose them.

When it does, the settings screen shows "A change is waiting" with today's version and the proposed
one side by side, and an owner approves or rejects. A member sees the pending change and a note
that an owner needs to approve it.

<Callout>
  That gate exists because a company instruction is the highest-leverage text in the workspace. Anyone
  who could get text in there could steer every future conversation for everyone, so the write is
  deliberately a human act.
</Callout>

Edit yours under **Settings → Instructions**, at either the workspace or the project level, and
your own under your profile.