> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mixus.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Instruction Placement Guide

> Learn where to put instructions for your AI agents for maximum effectiveness

Learn how to place instructions effectively across different layers of mixus to ensure your AI agents follow your guidance accurately and consistently.

## Overview

mixus has multiple layers where you can provide instructions to the AI. Choosing the right layer ensures your instructions have maximum impact and don't get lost in long conversations.

<Note>
  **Key Principle**: Put instructions as close as possible to where they'll be
  used. The closer the instruction is to the action, the more likely the AI will
  follow it.
</Note>

## Instruction Layers (Priority Order)

<AccordionGroup>
  <Accordion title="1. Organization Rules (Highest Scope)" icon="building">
    **Impact**: High | **Scope**: All users in organization

    Organization-wide rules that apply to every conversation and agent in your organization.

    **Best for:**

    * Company communication standards
    * Compliance requirements
    * Brand voice guidelines
    * Industry-specific terminology

    **How to set:**
    Go to **Settings** → **Organization** → **AI Rules**

    **Example:**

    ```
    Always use formal language in external communications.
    Never share proprietary pricing information.
    Always CC legal@company.com on contract discussions.
    ```
  </Accordion>

  {' '}

  <Accordion title="2. User Custom Instructions" icon="user">
    **Impact**: Medium-High | **Scope**: Individual user Personal preferences that
    apply to all your conversations. **Best for:** - Personal communication style

    * Signature preferences - Response format preferences - Language/tone
      preferences **How to set:** Go to **Settings** → **Chat** → **Custom
      Instructions** **Example:** `Keep responses concise and bullet-pointed. Use
        casual, friendly tone. My timezone is PST.`
  </Accordion>

  {' '}

  <Accordion title="3. Agent System Prompt" icon="robot">
    **Impact**: Medium-High | **Scope**: Specific agent Custom instructions that
    define how a specific agent template behaves. **Best for:** - Agent-specific
    behavior customization - Domain expertise injection - Workflow-specific rules

    * Output format requirements **How to set:** When creating/editing an agent,
      configure the **System Prompt** field. **Example:** `You are a sales report
        analyst. Focus on: - Quarter-over-quarter comparisons - Revenue trends by
        region - Top performing products Always format output as a markdown table.`
  </Accordion>

  {' '}

  <Accordion title="4. Agent Step Instructions" icon="list-check">
    **Impact**: Medium | **Scope**: Specific step Instructions for individual
    steps within an agent workflow. **Best for:** - Step-specific guidance - Task
    descriptions - Context for each action **How to set:** Define each step's
    description when creating the agent. **Example:** `Step 1: "Search my Gmail
          for emails from investors in the past week" Step 2: "Create a summary document
          with key points and action items" Step 3: "Draft a response addressing their
          main concerns"`
  </Accordion>

  <Accordion title="5. Memory Files" icon="brain">
    **Impact**: Medium | **Scope**: Retrieved when relevant

    Knowledge stored in your memory that the AI can retrieve.

    **Best for:**

    * Reference information
    * Templates and examples
    * Historical data
    * Domain knowledge

    **How to use:**
    Upload files to Memory or save chat content to memory. The AI retrieves relevant information automatically.

    **Example:**
    Save a file called "Email Templates.md" with approved response templates.
  </Accordion>
</AccordionGroup>

## Decision Guide: Where to Put Instructions

<CardGroup cols={2}>
  <Card title="Organization Rules" icon="building">
    Use when the instruction should apply to **everyone** in your organization,
    for **every** conversation. Examples: - "Always CC compliance on financial
    discussions" - "Use metric units for measurements"
  </Card>

  {' '}

  <Card title="User Instructions" icon="user">
    Use for **personal preferences** that should apply to all **your**
    conversations. Examples: - "I prefer bullet points" - "My timezone is EST"
  </Card>

  {' '}

  <Card title="Agent System Prompt" icon="robot">
    Use when the instruction should apply to **all runs** of a **specific agent**.
    Examples: - "Focus on Q4 metrics" - "Output in JSON format"
  </Card>

  <Card title="Agent Steps" icon="list-check">
    Use for **task-specific** instructions within a workflow. Examples: -
    "Search emails from last 7 days" - "Draft reply addressing pricing"
  </Card>
</CardGroup>

## Best Practices

### Do's

<Check>Write goal-oriented instructions, not tool-specific ones</Check>

```diff theme={null}
- "Use the sendEmail tool with to='john@company.com'"
+ "Send an email to John about the project update"
```

<Check>Use first-person language in agent steps</Check>

```diff theme={null}
- "Search the user's Gmail account"
+ "Search my Gmail account"
```

<Check>Keep instructions concise and specific</Check>

```diff theme={null}
- "When drafting emails, make sure to be professional and include a greeting
   and signature and proofread carefully and check for typos..."
+ "Draft professional emails with proper greeting and signature"
```

<Check>Put instructions where they're most relevant</Check>

```diff theme={null}
- Organization rule: "Always attach the Q4 report when discussing sales"
+ Agent step: "Attach the Q4 report to this email"
```

### Don'ts

<Warning>Don't repeat the same instruction in multiple layers</Warning>

If you say "be professional" in org rules, user instructions, AND agent prompt, you're wasting context space.

<Warning>Don't use overly technical language</Warning>

The AI understands natural language. Write instructions as if explaining to a colleague.

<Warning>Don't include implementation details in agent steps</Warning>

Let the AI figure out how to accomplish the goal. Focus on what, not how.

## Handling Long Conversations

<Tip>
  **Context Bloat**: In long conversations, earlier instructions may become less
  effective as the AI's context fills up. Here's how to manage this:
</Tip>

### When Context Gets Too Large

1. **Start a Fresh Chat**: For new topics, starting a new chat gives the AI a fresh context
2. **Use Agent Chaining**: Split complex workflows into multiple agents that pass results to each other
3. **Create Micro-Agents**: Build specialized agents for specific tasks rather than one agent that does everything

### Signs of Context Bloat

* AI stops following earlier instructions
* AI uses wrong tools or approaches
* Inconsistent behavior during long conversations
* AI "forgets" information from earlier in the conversation

## Practical Examples

### Example 1: Sales Report Workflow

<CodeGroup>
  ```text Organization Rules theme={null}
  All financial reports must include disclaimer about forward-looking statements.
  Use fiscal year definitions (July 1 - June 30).
  ```

  ```text Agent System Prompt theme={null}
  You are a sales analyst specializing in B2B SaaS metrics.
  Focus on: ARR, MRR, churn rate, and customer acquisition cost.
  Present findings in executive summary format.
  ```

  ```text Agent Step 1 theme={null}
  Retrieve the latest sales data from my Salesforce dashboard
  for the current quarter.
  ```

  ```text Agent Step 2 theme={null}
  Create an executive summary comparing this quarter to last quarter,
  highlighting significant changes in key metrics.
  ```
</CodeGroup>

### Example 2: Email Response Workflow

<CodeGroup>
  ```text User Instructions theme={null}
  Keep email responses under 200 words.
  Use my standard signature: "Best, Sarah"
  ```

  ```text Agent System Prompt theme={null}
  You are an email assistant for customer support.
  Tone: Helpful, professional, empathetic.
  Always acknowledge the customer's concern before providing solutions.
  ```

  ```text Agent Step theme={null}
  Draft a response to the customer's refund request,
  following our standard refund policy from memory.
  ```
</CodeGroup>

## Related Resources

<CardGroup>
  <Card title="Creating Agents" icon="robot" href="/agents/creating">
    Learn how to create and configure AI agents
  </Card>

  <Card title="Micro-Agent Patterns" icon="link" href="/agents/micro-agent-concepts">
    Advanced patterns for complex workflows
  </Card>

  <Card title="Agent Examples" icon="lightbulb" href="/agents/examples">
    Real-world agent configurations
  </Card>
</CardGroup>