> ## Documentation Index
> Fetch the complete documentation index at: https://docs.stardeck.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Guidance cards
> Break a customer agent's rules into small, reusable cards you can scope to a channel and toggle on or off—without rewriting the persona.
A **guidance card** is one rule the agent should follow, written as a small, self-contained card instead of buried in the persona. Cards keep the persona short and make each rule something you can scope to a channel, reorder, and switch off on its own.
A card is built from two parts:
* **Condition** *(optional)* — when the rule applies: "a customer sounds upset", "someone asks for a price".
* **Instruction** *(required)* — what to do: "reply gently and use their first name", "point them to the pricing page instead of quoting a number".
Together they read as **"If *condition*, *instruction*"**—or just the instruction, when there's no condition.
## Two kinds of card
Each card shapes one of two things, set by its **surface**:
* **Conventions** (reply surface) — shape *how the agent replies*. Tone, formatting, what to say or avoid, where to send people.
* **Scoped rules** (classifier surface) — shape *whether the agent replies at all*. These feed the [reply classifier](/customer-agents/operator-workflow#the-reply-classifier): "don't reply to thanks-only messages", "always reply when someone mentions a refund".
The editor's surface toggle switches between the two.
## Categories
Conventions are grouped into a fixed set of categories. The grouping organizes the editor and the order the rules reach the agent—framing first, catch-all last:
| Category | For rules about… |
| --------------------- | ------------------------------------------ |
| **Voice & tone** | How the agent sounds. |
| **Clarification** | When and how to ask the customer for more. |
| **Content & sources** | What to cite, link, or refuse to answer. |
| **Safety** | Hard limits and sensitive topics. |
| **Policies & other** | Anything else. |
Scoped rules (classifier cards) aren't categorized—they're a flat list.
## Scope a card to a channel
By default a card applies on **every** channel. Give it a **channel scope** to narrow it—for example, a "never quote prices" rule only on LINE, or a warmer greeting only on Instagram. A card with no channel scope applies everywhere; a scoped card applies only on the channels you pick.
Channel scope is the only targeting available today. A card is either global or limited to specific
channels—there's no per-customer targeting yet.
## Order and on/off
* **Reorder** cards within a category to control the order the agent reads them.
* **Toggle** any card off to retire a rule without deleting it—useful while you tune behavior.
## How cards reach the agent
At reply time the agent collects every enabled card whose scope matches the live conversation, then:
* **Conventions** are grouped by category and added to the agent's instructions as a `Conventions` section, right after the persona.
* **Scoped rules** are added to the classifier's reply-or-stay-quiet decision.
Because cards are scoped and ordered deterministically, the same conversation always assembles the same instructions—so they stay cheap to run.
## When to use a card vs. the persona
| Put it in… | When it's… |
| ------------------- | ------------------------------------------------------------------------------------------------------------- |
| **The persona** | Identity and voice—who the agent is, the through-line of how it speaks. |
| **A guidance card** | A specific, repeatable rule, especially one that only applies on some channels or that you'll want to toggle. |
Keeping rules in cards means the persona stays a short, readable brief instead of a wall of edge cases.
## Next steps
Add a goal and a handoff on top of conventions
How an agent decides to stay quiet
Where cards fit in the workspace