PromptZone - AI Prompts, Guides and Tools for Builders

Cover image for Stop re-pasting your resume into Claude: give it state over MCP
Amine Daddou
Amine Daddou

Posted on

Stop re-pasting your resume into Claude: give it state over MCP

Prompting Claude for a job search goes well for about one application. Then things start to slip:

  • the resume gets pasted again in every new chat;
  • the tailored versions end up spread across conversations you can't search;
  • "which roles did I apply to, and with which resume?" lives in a spreadsheet Claude never sees.

None of that is a prompting problem. Claude writes a good cover letter. What it doesn't have is state: nothing carries over from one job to the next unless you paste it in again.

Better prompts don't fix it

The usual workarounds are all prompt-side:

  • a long system prompt or project file with your whole history in it;
  • a "master resume" you paste before every request;
  • asking Claude to "remember" roles it has no way to store.

Each one makes the context window do a database's job. The context gets longer, the instructions compete with each other, and the one thing you need (what's already been sent, and where) still isn't recorded anywhere.

Move the state out of the prompt

The alternative is to keep the prompt short and give the model tools that read and write durable state. With the Model Context Protocol (MCP), that state can live in a small service Claude calls, instead of in the conversation.

For a job search, that state is three things:

  1. Your profile, as structured data. We use JSON Resume, so the model edits fields, not a blob of prose.
  2. The board: one record per job, with its status.
  3. The documents: a tailored resume and cover letter per job, rendered as PDFs, so the output is a file, not a chat message you copy out.

With that in place, the prompt goes from "here is my entire career, now write X" to "file this role and draft the resume for it". The model fetches what it needs through the tools.

What changes in practice

  • Shorter prompts, fewer contradictions. Tool descriptions carry the rules, not a growing system prompt.
  • Edits stick. Changing your profile once changes every document rendered after it.
  • You can audit it. What Claude filed and generated is on a board you can open, not buried in chat history.

[Add one or two things you learned while building it: a tool description you had to rewrite, a failure mode you hit, a real before/after.]

What we're building

This is the idea behind SuperApplicant: a personal job board that Claude runs over MCP. Claude files the jobs you find, and SuperApplicant renders a tailored resume and cover letter for each one as PDFs. It runs on the Claude you already pay for, not a second AI subscription.

It isn't open yet. If you want in when it is, there's an early access list. The free plan will include 3 rendered documents.

I'd like to hear how you handle state between Claude sessions today. Projects, memory files, a custom MCP server? What has held up?

Top comments (0)