The difference in a nutshell

Think of a new collaborator. Custom instructions are the rulebook you hand them on day one: "always answer in formal Italian, you're a legal assistant, never give definitive opinions." They orient them, but give them access to nothing: they're words, fixed instructions.

MCP, instead, is handing them the keys to the archive and access to the management software: now they can go and fetch real documents and carry out operations. One defines the attitude, the other provides the tools. A collaborator with only rules and no access knows how to behave but can't touch anything; one with access and no rules can do damage with the wrong tone.

Which to choose for your case

  • Custom instructions when you want the AI to respond a certain way, always: a tone, a role, constraints, a format. No external data is needed, what's needed is consistency in behavior. It's the simplest and lightest choice.
  • MCP when the task requires data that changes or actions in the real world: reading your up-to-date documents, checking fresh information, writing to a file. Instructions alone aren't enough here, because the AI would have nothing to access.
  • Together when you want both: the AI behaves as you say (instructions) AND works on your real data (MCP).

A concrete example

Anna is a lawyer and uses the AI for first drafts. With custom instructions she gives it the fixed behavior: "write in clear legal Italian, always structure with premises and conclusions, never state a rule unless I've provided it." This applies to every conversation, without repeating it.

But she also needs the AI to work on her real filings. Here instructions aren't enough: she connects a filesystem MCP server to the folder of templates. Now the AI behaves as the lawyer wants (thanks to the instructions) and draws on the real documents (thanks to MCP). Anna has understood the difference: the rules shape the how, MCP provides the what. She needed both.

When one isn't enough (and how to fix it)

If you gave instructions but the AI doesn't "see" your data

Custom instructions give access to nothing: if you ask the AI to work on your files and it answers empty, it's because the connection is missing. Add an MCP server to give real access. No instruction, however detailed, makes data appear that the AI can't reach.

If you connected MCP but the AI behaves randomly

If the AI accesses the data but responds with the wrong tone or format, the rules are missing. Add custom instructions to set behavior and style. MCP gives the tools, not the good manners.

If you repeat the same instructions in every chat

If you find yourself re-pasting the same rules every time, you're not using custom instructions: those should be set once in the assistant's settings and apply always. Look in the settings for the custom instructions section and put them there.

A tip from someone who actually uses it

Always start from custom instructions: fixing how the AI behaves is free, immediate, and improves every answer, even without connecting anything. Add MCP only when you realize you need to let the AI access something real. Many people set up complex connections when their real problem was just giving the AI three lines of clear rules.

Frequently asked questions

Are custom instructions a kind of memory?

No, and it's a common confusion. Instructions are fixed rules that apply the same way in every chat; they don't "remember" what happened in conversations. If you want the AI to remember facts that change over time (your projects, the decisions made), you need a memory, which can also be provided via MCP with a dedicated server.

Can I put my rules inside an MCP server instead of in the instructions?

Technically some things overlap, but it's a waste: behavior is set better with custom instructions, which are free and lightweight. Reserve MCP for what instructions can't do, namely access data and perform actions. Using the right tool for each purpose keeps everything simpler.

If I use MCP, can I do without instructions?

Here's the misunderstanding to clear up: no, because they solve different problems. MCP gives the AI the power to access and act, but doesn't tell it by what criteria. Without instructions, an AI with many accesses is a capable collaborator but with no rules. The two together — clear rules plus the right accesses — are what makes the assistant truly reliable.