For the Flip the Switch Hackathon I built a small tool to stop me doing that. It's called ProofPack.

Demo video: https://youtu.be/-XioOO6ABtM
Code: https://github.com/Thecynic007/flip-the-switch-proofpack

The idea

Most AI tools will happily write you an impressive CV. I wanted the opposite. ProofPack asks one question: what evidence do you actually have?

You paste in a job description and notes about your real projects. Three AI agents then work through it together in one shared room:

• scout reads the job and lists the key requirements
• mapper matches each requirement to your evidence and marks it supported, partly supported or missing
• checker flags claims you can't back up, suggests honest wording, adds a 7-day next step for each gap and writes five interview questions about your own work

Then checker asks me to approve or revise, and saves the result as a file.

Why Switch

Switch is a shared room where people and AI agents work together. What made this project possible is that agents can hand work to each other by mentioning each other by name. Scout finishes and tags mapper, mapper finishes and tags checker, and I can read the whole conversation in one place and step in whenever I want.

Without that, I would be copying and pasting between three separate chats and losing track of who said what. The room, the shared history and the handoffs are the whole product, which is why I chose Switch rather than a single chatbot.

How I set it up

I built this on a Windows laptop. You need Docker Desktop, Switch Console, Node.js and an agent provider. I used the Codex CLI (version 0.167.0), signed in with a ChatGPT account, and checked it with codex --version.

In Switch Console I started a local server, created three agents called scout, mapper and checker, and gave each one a short set of instructions. I created one room and added all three agents to it.

One setting matters a lot. "Who can talk to your agent" has to allow other agents, or they won't take work from each other.

What went wrong

I started with OpenCode and an NVIDIA model. Switch Console kept showing a sign-in warning and wouldn't let me select OpenCode when creating an agent. I tried several fixes, then switched to Codex, and the problem went away. It taught me that when a tool's own warning won't clear, the faster move is sometimes to test a different path instead of fixing the first one.

I also learned that checker waits for my approve or revise reply in the same thread before it saves anything. For a while I was waiting for a file that was waiting for me.

The result

I tested it on a made-up applicant called Sam, with two small projects and a couple of courses. For a junior backend job, the result was: 0 requirements supported, 3 partly supported, 3 missing.

Here is one row from the table checker produced:

I also ran a junior frontend job with the same applicant and got 0 supported, 3 partly supported and 2 missing. I tested the revise step too: I asked for shorter wording, mapper rewrote it, checker rechecked it, and I approved the new version.

What I learned

The most useful part wasn't the table. It was the interview questions. Reading a question like "which parts of this project did you write yourself?" was uncomfortable, because a lot of my own project work was done with AI help.

I used AI assistants to help plan and build ProofPack, and I can explain how each part works. I also used Wispr Flow to speak many of my questions and prompts while planning and debugging, and I typed the terminal commands and edited the files myself.

Limits

The AI can get things wrong or even invent evidence. That is exactly why checker flags what isn't backed up and why I make the final decision. ProofPack doesn't write your CV, scan GitHub or scrape job sites. All the sample data is made up. It's a hackathon prototype, not a hiring tool.

What I'd do next

I'd let people paste in their own project notes and keep a history of past runs, so you can watch your gaps shrink over time. I would also test it on more kinds of jobs than backend and frontend.

The code, sample data and setup steps are all in the repo if you want to try it.