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

# Skills & LLM Instructions

> Push always-on system instructions to the AI client on every connection — steer the model toward the right tools without writing a prompt.

## Overview

**Skills** are server-level instructions sent to the AI client automatically on every MCP connection. They appear in the `instructions` field of the MCP `initialize` response and most AI clients (Claude Desktop, Claude Code, Cursor, ChatGPT, VS Code) pass them verbatim into the LLM system prompt.

Skills let you steer the model's behaviour — which tools to prefer, in what order to call them, what assumptions to avoid — without the user needing to type anything.

***

## Where to Configure Skills

Open **GetMCP → Servers → \{Your Server} → Skills tab**.

Write your instructions in the text area and click **Save Skills**. The instructions take effect immediately on the next MCP `initialize` call — no restart required.

The field accepts plain text only, up to **8,192 characters**.

<img src="https://mintlify.s3.us-west-1.amazonaws.com/infiwebs/images/skills-editor.png" alt="Server — Skills tab with instructions text area and Save Skills button" />

***

## Skills vs Prompts

GetMCP has two related but distinct features for shaping LLM behaviour:

|                          | Skills                                     | Prompts                                |
| ------------------------ | ------------------------------------------ | -------------------------------------- |
| **Delivery**             | Pushed automatically on every `initialize` | Pulled on demand by the user           |
| **When applied**         | Always — every session                     | When the user selects a named template |
| **Configured in**        | Server → Skills tab                        | Server → Prompts tab                   |
| **MCP field**            | `initialize.instructions`                  | `prompts/get` response                 |
| **User action required** | None                                       | User must invoke the prompt by name    |

Use **Skills** for permanent, always-on guidance. Use **Prompts** for optional named templates the user can pick from a list.

***

## What to Write

Good Skills instructions are **specific and action-oriented**. Vague instructions ("be helpful") are usually worse than no instructions at all.

### Effective patterns

**Enforce a call order:**

```
When the user asks about an order:
1. Always call lookup_order first to fetch the current state.
2. Never suggest a refund without calling verify_payment_status first.
3. Use search_products for inventory questions — do not guess stock levels.
```

**Steer away from wrong tools:**

```
This server is connected to our internal CRM only.
Do not use web_search or any external data source for customer information.
Always use get_customer and list_interactions for any question about a contact.
```

**Set data-handling expectations:**

```
All monetary values are in USD cents (integer). Divide by 100 before displaying.
Dates are ISO 8601 UTC. Always convert to the user's local timezone before showing.
```

**Reduce unnecessary calls:**

```
list_products is expensive. Call it once per session and cache the result.
Only call refresh_inventory if the user explicitly asks for updated stock data.
```

### What to avoid

| Anti-pattern                                   | Why                                                                                                    |
| ---------------------------------------------- | ------------------------------------------------------------------------------------------------------ |
| Repeating the tool's description               | Already in `inputSchema.description` — duplication adds tokens, not value                              |
| Generic politeness instructions                | MCP Skills compete with the AI client's own system prompt; generic tone guidance is usually overridden |
| Instructions that contradict the tool's schema | The tool schema is authoritative; conflicting instructions confuse the model                           |
| Very long instructions (> 2,000 chars)         | Longer isn't always better — token budget is shared with the conversation                              |

***

## How the AI Receives Skills

GetMCP returns the Skills text in the MCP `initialize` response:

```json theme={null}
{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "protocolVersion": "2025-11-25",
    "serverInfo": { "name": "My Server", "version": "1.0.0" },
    "capabilities": { "tools": {} },
    "instructions": "When the user asks about an order:\n1. Always call lookup_order first..."
  }
}
```

Each AI client handles this field differently:

| Client                           | Behaviour                           |
| -------------------------------- | ----------------------------------- |
| **Claude Desktop / Claude Code** | Injects into the system prompt      |
| **Cursor**                       | Injects into the system prompt      |
| **VS Code (Copilot)**            | Injects into the system prompt      |
| **ChatGPT**                      | Injects into the model instructions |
| **Custom clients**               | Depends on implementation           |

<Note>
  If a client ignores the `instructions` field, Skills have no effect. Check the client's documentation if you notice instructions not being followed.
</Note>

***

## Skills and Multiple Servers

Each server has its own independent Skills. When an AI client connects to multiple servers simultaneously, each server's instructions are injected separately. Keep instructions scoped to the tools on that server to avoid confusion.

***

## Character Limit

The Skills field accepts up to **8,192 characters** (\~1,500–2,000 tokens). This is intentionally constrained — overly long instructions consume token budget that should be available for the actual conversation.

If you find yourself hitting the limit, consider:

1. Moving static reference data (product lists, IDs) into a **Resource** that the AI can fetch on demand.
2. Using a **Prompt** for optional context the user can explicitly load.
3. Trimming instructions to the highest-value, most-specific guidance.