---
title: Assess candidates on your own repo
description: Test candidates on a repository that closely resembles the job they'll do at your organisation.
url: https://cmdround-docs.thally.app/guides/assess-on-your-own-repo
---

# Assess candidates on your own repo

Test candidates on a repository that closely resembles the job they'll do at your organisation.

Custom repositories have always been how engineers were tested before agentic
development. CmdRound takes note of this and has this exact feature. You can test
candidates on a repository that closely resembles the day-to-day job they'll be
responsible for at your organisation.

## How a task works

A task is a ticket on your own repository plus the checks that score it. A task contains
four parts. Let's take a look.

> **Tip:**
  If writing each task by hand looks like a long process, you'll be happy to know we have
  "Write it with your coding agent". It gives you a prompt for an agent in your codebase,
  and the agent gives back a JSON block we use to fill in every part.
  [See how it works](#write-a-task-with-your-coding-agent).

![The "Write it with your coding agent" box with the "Copy prompt" and "Paste answer" buttons](/images/agent-box.png)

Back to the four parts of a custom task in your repository:

- The ticket. This is what the agent your candidate drives sees. It goes into the repo as
  `TICKET.md`.
- The note for the candidate. The twist here is that only the candidate sees this. It
  tests how well a candidate can read an instruction and pass it to the agent. If they
  copied and pasted this note directly to the agent, you will be notified in the replay
  section of that assessment.
- Checks. Each check is a command. If it exits with `0`, it passed. A check works with
  the ticket we spoke of earlier to make sure the work listed there is done.
- Hidden test files. These files aren't added until the round ends, and the candidate
  never sees them. The checks run them.

## Add your repo

Click "Your repos" at the top of your dashboard, or visit
[cmdround.com/dashboard/repos](https://cmdround.com/dashboard/repos). Paste the link to
the repository and click "Add repo". The repository has to be public, and you can add up
to three.

![Your repos page with no repos yet](/images/add-repo.png)

After you click "Add repo", a setup process starts. The repo is set up once, so every
candidate gets the same copy in their own isolated workspace. The setup typically takes a
minute or two.

![A repo waiting to set up](/images/repo-waiting.png)

![A repo setting up](/images/repo-setting-up.png)

When the setup is done, its status changes to "Ready" and you can add your tasks, by
hand or, as we'll see soon, with an agent.

![A repo marked Ready](/images/repo-ready.png)

## Write a task with your coding agent

We advise that you have an agent that lives in your codebase write the task. We provide
it with enough context about what CmdRound is and how our tasks, assessments and checks
work. This helps it write the best possible and fairest assessment for your candidates.

To get started, open the repo you just added and click "New task". Then click
"Copy prompt" in the "Write it with your coding agent" section.

![The New task page with the "Write it with your coding agent" section at the top](/images/new-task.png)

Paste the prompt into any agent that has full context of how your codebase works. Your
agent will ask you questions like these:

![A coding agent asking about the company, role, seniority and skills before it writes the task](/images/agent-questions.png)

After your detailed answers, the agent writes all parts of the task: the ticket, the
note, the checks and the hidden test files. Then it runs every check three times: on the
untouched repo, on a correct solution and on a solution that ignores the note. Only after
every check behaves as it should does it give you its answer, with the whole task as a
JSON block at the end.

Copy the agent's whole answer, click "Paste answer", paste it into the box and click
"Fill the form". We read the JSON and fill in the form on your behalf. Read through each
field to be sure the agent did not stray from your goal.

![The paste box open under "Write it with your coding agent"](/images/paste-box.png)

![The task form filled in from the agent's answer](/images/task-form-filled.png)

CmdRound only accepts the agent's work if the runs prove the task works and can't be
gamed by a curious candidate. If any check fails, we tell you which check failed and
leave the form empty.

## Save the task

When the checks have passed and the task meets your goal, click "Save task". It will
then appear in your tasks when you create a new assessment.

## Try a dry run

To check the integrity of your task, we offer a free dry run. Click "Start the dry run".
In the background, we give an agent the ticket and let it work alone, with no human in
the middle. It takes a few minutes, depending on the size of the ticket and the
repository.

At the end of the dry run, a good task shows "Agent alone failed": the agent could not
fix the bug or build the feature alone, and it needs a human in the middle. This means
the task is good for assessing your candidates.

![A dry run verdict on a saved task](/images/dry-run-verdict.png)

## Put the task on an assessment

To put your task to work, click "New assessment" on your dashboard. Make an assessment
the same way as in [Send your first assessment](/quickstart#steps). On the tasks step,
your new task is listed in the "Your tasks" section. Pick it and carry on as usual.

![The tasks step with your own task under "Your tasks"](/images/builder-your-tasks.png)

## Read the result

When candidates submit their work, you are notified, and this is the part where it all
comes together. The scorecard in the replay, which you can open right from the mail or
from each candidate's card on your dashboard, answers four questions:

1. Do they verify the AI's work?
2. Does the work hold up?
3. How do they recover when it breaks?
4. How do they direct the agent?

Each one gets Standout, Strong, Mixed or Flag.

  ![The scorecard at the top of a replay](/images/sample-scorecard.png)

Here is a sample replay. It shows key moments, the candidate's prompts and the
permissions they gave the agent.

  ![Key moments in a replay](/images/sample-moments.png)

  ![How the round ended](/images/sample-how-it-ended.png)

## Troubleshooting

### The repo setup failed

If the setup of your repository doesn't work, you see what your repo printed last in the
setup process. Fix the command, then click "Set it up again" to run it again.

![A failed repo setup with its last output and the "Set it up again" button](/images/repo-failed.png)

### The agent's answer was refused

A task can also be refused when a check fails. Copy the error message, give it back to
your agent and paste its new answer.

![A pasted answer refused because a check failed](/images/paste-refused.png)