AI toolsGoogle AI · 3/342 seconds
Google Jules: delegating a coding task
One scoped task, a GitHub repository, a test command. Jules proposes a plan, executes, and hands back code to review.
In short
One task is handed to Jules: add reading progress to a small dependency-free app.
Jules clones the repository into a virtual machine, proposes a plan, executes, runs the project's tests and returns a complete diff.
What you'll learn
- A delegated task is scoped by three things: a precise goal, constraints, a test command.
- The plan is proposed before any change: that is where you keep control.
- Jules executes a scoped task and hands back code to check. It does not replace review.
Step 1 / 5
Connect the repository
GitHub gets connected, then the repository gets picked. Here, a dependency-free reading-list app: HTML, CSS and browser JavaScript.
The repository carries a working agreement that sets the rules: keep the app dependency-free, put testable logic in a dedicated module, preserve the visual direction, and run the tests before completing.
Step 2 / 5
Scope the task
The request describes a precise goal, the constraints to respect and the test command to run.
The more explicit the verification command, the less the review is about execution details.
Step 3 / 5
Approve the plan before the code
Jules clones the project into a virtual machine and proposes a plan before changing anything.
The steps get reviewed, then execution starts.
Step 4 / 5
Execution and diff
The work happens in the background: Jules edits the necessary files and runs the project's tests.
At the end, the summary and the full diff read as one block.
Step 5 / 5
Fix, then publish
If something is off, a correction can be requested straight away, on the same thread.
Once the diff is accepted, a branch or a pull request gets published.
What came out
- Reading progress added to the demo app, with progress stored locally.
- The project's tests executed by Jules in its own environment before handing back.
- A complete, readable diff before any branch is published.
What went wrong
- The output is still code to review: nothing is merged without a human decision.
- The episode shows one isolated task on a dependency-free demo repository. A denser real-world project is not tested here.
Sources
- https://jules.google/docs/
https://jules.google/docs/
- https://jules.google/docs/running-tasks/
https://jules.google/docs/running-tasks/
- https://jules.google/docs/code/
https://jules.google/docs/code/
Behind the scenes
This episode came out of a pipeline, not twelve browser tabs.
Research, script, shots, media, editing, checks and publishing kit: ARGO Studio holds everything that produced this video.