Allow to provide further instructions to the LLM from an MCP server response

I’m using the sonnet model. All screen below are done with the agent.

I have this tool

const svelteDocsTool = {
	name: 'get-svelte-docs',
	description:
		'Read the complete Svelte documentation. Use this for any request related to Svelte.',
	async processInput(_input: unknown): Promise<string> {
		try {
			const response = await fetch('https://svelte.dev/llms-full.txt');
			if (!response.ok) {
				throw new Error(`Failed to fetch docs: ${response.statusText}`);
			}
			const docsContent = await response.text();

			// Write the docs to file
			const filePath = join(projectRoot, '.personal', 'svelte-5-docs.txt');
			await writeFile(filePath, docsContent, 'utf-8');

			return `The Svelte documentation has been saved. Please read it from ".personal/svelte-5-docs.txt". You must read the file, it's an order.`;
		} catch (error) {
			if (error instanceof Error) {
				throw new Error(`Failed to fetch or save Svelte docs: ${error.message}`);
			}
			throw error;
		}
	}
};

Cursor ignores the response that tells it that needs to read it from a file and just gives me outdated info:

Cursor even claims that it doesn’t have the ability to read it unless i tell it that it actually can like in the screen above


That txt file is reachable from cursor, unfortunately i can’t prove it because the list of included files is currently broken

Hey, cool idea! Can you try to explicitly tell the Composer to use the response from the tool as instructions on what to do next? If that works, you may be able to use the new Project Rules to enforce this long term.

I have one tool that’s actively editing one of the .mdc files inside .cursor/rules