> For the complete documentation index, see [llms.txt](https://doc.aissist.io/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://doc.aissist.io/capabilities/instructions-and-information.md).

# Instructions and Information

How Aissist's AI decides what to say - instructions tell it what to do, information tells it what it knows.

*Last updated: October 25, 2026*

Everything the AI does in a conversation comes from two things you give it: **instructions** (what to do) and **information** (what it knows). This page explains both, the levels they are set at, and how they combine in each reply.

### Instructions: what to do

Instructions shape behavior. Use them for:

* tone, style and reply format
* rules and limits, such as "never confirm a delivery date"
* workflows, step by step, such as how to handle a return
* when to hand over to your team

### Information: what it knows

Information is the facts the AI answers from:

* **Context** you write, such as opening hours or what you sell
* **Assets** it searches, such as your help center, policy pages and documents (see [Assets](https://doc.aissist.io/capabilities/assets.md))
* **Action results** it reads from your systems, such as an order's status (see [Actions](https://doc.aissist.io/capabilities/actions.md))

Instructions say how to use the facts. Information supplies the facts. Keep them apart: a return policy is information and belongs in an asset; "ask for the order number before discussing a return" is an instruction.

### The three levels

| Level                   | Where you set it                                                                                            | When it applies                                                 |
| ----------------------- | ----------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------- |
| **Workspace**           | **Workspace** tabs: **Character**, **Instruction**, **Context**, **Greeting**, **Escalation**, **Advanced** | Every conversation                                              |
| **Sub-agent (session)** | A sub-agent's **Instruction**, plus **Context**, **Task** and **Summary** under **Options**                 | Once its scenario is detected, for the rest of the conversation |
| **Message**             | Assets, action results and continuous learn examples                                                        | Only when they match the current message                        |

**Workspace** is for what is always true: who the agent is (**Business Name and Description**, **Identity**, **Purpose**, **Style** on the **Character** tab), universal rules, shared business facts, and global escalation rules. Keep it short, because it is in every reply.

**Sub-agent** is for one scenario, such as returns or order tracking. Its instructions and context join only when that scenario is detected, so they do not crowd other conversations. See [Sub Agents](https://doc.aissist.io/capabilities/sub-agents.md).

**Message** level is retrieved fresh for each message: the asset passages that match the question, the results of any actions called, and examples you taught through [Continuous Improvement](https://doc.aissist.io/capabilities/continuous-improvement.md) for conversations like this one.

### How they combine

For each reply, Aissist puts together the workspace instructions and context, the instructions and context of the sub-agents active in the conversation, and the information retrieved for this message. The reply is written from all of it.

To see what went into a reply, open the conversation under **Deploy → Sessions**. Each AI reply lists its **Sources**.

So:

* **Conflicts cause errors.** If a workspace instruction says one thing and a sub-agent or asset says another, the reply can follow either. Remove the conflict instead of hoping one wins.
* **Noise lowers accuracy.** Rules that only matter for one scenario belong in that sub-agent, not the workspace.

### With Ask Engine

Ask Engine can explain why the agent behaved the way it did, and what changed:

```
Why did the AI offer a refund in conversation 48213? Which sub-agents and sources did it use?
```

```
What changed in this workspace's instructions or sub-agents last week?
```

Ask Engine reads the conversation, the sub agents and the change history, and answers with what it found. It does not edit instructions; to teach a better answer, ask it to add a continuous learn example, which it previews and adds only after you confirm. See [Ask Engine](https://doc.aissist.io/capabilities/ask.md).

### Best practice

* **Start with one clear objective.** "Give concise pricing guidance. Do not quote exact prices without quantity and ZIP code. Hand over when pricing depends on data you do not have" beats "Answer pricing questions."
* **Say what not to do.** The AI tries to be helpful and fills gaps. Add explicit limits: do not guess, do not estimate unavailable values, hand over when unsure.
* **Write workflows as numbered steps.** Start a sub-agent instruction with a short workflow title, then list the steps in order. Ask one thing at a time.
* **Prefer intent over exact wording.** "Confirm the customer can order a single item" works better than a fixed sentence. Use exact wording only for legal text or approved promises.
* **Keep instructions bounded.** Replace "explain thoroughly" or "if relevant" with limits such as "under 3 sentences" or "one question at a time".
* **Define handover triggers exactly.** Not "escalate if necessary", but "hand over when the customer asks for a person, or after the order number, reason and photos are collected".
* **Prefer lists over tables** in instructions.
* **Put each thing in its place:** universal rules in **Instruction**, shared facts in **Context**, scenario logic in sub-agents, reference knowledge in assets.

### Related

* [Create a Workspace](https://doc.aissist.io/tutorial/create-workspace.md)
* [Sub Agents](https://doc.aissist.io/capabilities/sub-agents.md)
* [Assets](https://doc.aissist.io/capabilities/assets.md)
* [Try](https://doc.aissist.io/tutorial/try.md)
