AI & MCP

Prompts for agents, not chat: contract, sequence and stop

A prompt for an agent is an operating contract: tools, order, format and stop rules. Tone comes later, and often it does not come at all.

Key points

  • A chat prompt describes a tone; an agent prompt describes a verifiable process.
  • The tool sequence must be written down: without order the model improvises calls.
  • The prompt lives in a versioned file, so a change reads as a diff.

A prompt for an agent is an operating contract: it states which tools exist, in what order to use them, what counts as a valid result and when to stop. Tone comes later, and often it does not come at all.

Why chat is not enough

In a chat the answer is the product. In an agent the answer is an intermediate step: the model calls a tool, reads the result and decides. The prompt must therefore describe what counts as evidence, not only how to sound.

Four blocks

  • Scope: what the agent may touch and what it must refuse, with permissions named explicitly.
  • Sequence: the order of tools, for example design system, search, read, compose.
  • Format: the shape of the expected output, including required fields such as URL and command.
  • Stop: the conditions for stopping and asking, instead of trying another variant.

The sequence belongs in the prompt

EneUI publishes the sequence in /AGENT.md: get_design_system, then search_components or search_blocks, then get_component, then compose_page, and get_page when prose is needed. Written that way, the choice does not depend on the model's mood.

curl -s https://ui.eneawork.it/agent-skill.md | head -40

The file is read once and stays in context. It beats a prompt pasted into a config, because the source is the same tool list the server exposes: when a tool is renamed, the text follows.

The prompt lives in the repository

A prompt in a versioned file can be reviewed, compared and restored. A prompt in a vendor panel cannot. The difference shows at the first failure: with the file you read the diff between the version that worked and the broken one.

Measure instead of hoping

Every call to the EneUI server writes an mcp_call row with the tool name and the plan, so the failed-call rate per tool is observable. Judge a prompt on that number, not on an impression after three tries.

The catalogue

One component an agent can propose is Jahed AI prompt box; the rest sit among 111,985 components and 1,556 blocks. The endpoint is on /mcp and the agent-side flow on /agents.

The limitation

A prompt is not a permission. The initialize instructions are two sentences and the model can ignore them; the server cannot enforce call order, because it does not see the conversation. The only guarantee sits in the client that limits tools and commands.

Where to continue

The file shape is in anatomy of an agent skill, the context is in what an AI agent is and the tools are in designing an MCP tool. For execution risks read securing a coding agent.

Frequently asked questions

What is the difference between a chat prompt and an agent prompt?

The first describes an answer, the second a process. An agent prompt lists tools, call order, output format and stop conditions, because the answer is only an intermediate step.

Should the tool sequence be written down?

Yes. Without an order the model tries random combinations and wastes calls. EneUI writes the sequence in /AGENT.md: design system, search, read, compose, then page.

Where should the system prompt live?

In a versioned file in the repository, so every change is a readable diff. A prompt in an external panel cannot be compared or restored as easily.

How do you measure whether a prompt works?

With the failed-call rate per tool and the number of steps needed to finish a task. The EneUI server records one mcp_call row per call, with the tool name and the plan.

Can a prompt block a dangerous command?

No. A prompt is text and the model can ignore it; the block lives in the client permissions. The server does not see the conversation and cannot stop a command run elsewhere.