How to use the Dogfood Branch QA agent skill (ce-dogfood)

What does the ce-dogfood skill make the agent do?

An agent skill is a folder of instructions and supporting files that an AI coding agent loads when you ask it to perform a specific job. ce-dogfood gives the agent a QA routine for a branch. It maps the user flows affected by your branch's changes versus the trunk, runs those flows in a real browser, and judges both whether they work and how they feel to the personas your product defines.

When it finds a small, low-risk breakage, it can fix it, add a regression test, and commit the change. Bigger decisions, such as anything needing an architectural or schema choice, a change to product behavior, or a change spanning many files, are not implemented. They are listed in the report under decisions for a human.

• Scope is the branch diff, not the whole app.
• Fixes are small and low-risk only, each with a regression test that fails before the fix and passes after it, or a stated reason why no automated test was meaningful.
• The output is a durable report saved in your repo.

How do you install the ce-dogfood skill?

Download the skill ZIP from the skill's page on this site and unzip it. Where it goes depends on how you want to use it.

For Claude Code, unzip it into ~/.claude/skills/ to make it available in all your projects, or into a project's .claude/skills/ folder to make it available only in that repository.

In the Claude apps, you can upload the ZIP instead under Settings, then Capabilities, then Skills.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside the repository

How do you start a dogfood run?

The skill is manual only. It will not start because you mention testing in passing; you invoke it by name. In Claude Code, type /ce-dogfood. You can add an argument to choose the target: a PR number, a branch name, or nothing to use the branch you have checked out. A --port PORT option is also accepted.

A PR number or a named branch is isolated in its own worktree through the ce-worktree skill, so your primary checkout is not switched. A blank target runs in place on the current branch.

• Blank argument: test the current branch in place.
• PR number: test that PR in an isolated worktree.
• Branch name: test that branch in an isolated worktree.

What does a run look like from start to finish?

The skill follows fixed phases: scope, analyze the diff, map the flows, derive a test matrix, start the local dev server, execute scenarios, run the fix loop, then write the report. Each scenario is judged one at a time for correctness and for how it feels to each persona.

A report file is created as soon as the test matrix exists, with every scenario marked Pending, and it is updated after each scenario. If a run is interrupted, the report on disk still shows where it stopped. The report is saved under dogfood-reports/ in your repository, named with the date and branch, for example as a file ending in -dogfood.md. If your repo sets docs_root in .compound-engineering/config.yaml, that folder is used as the root instead of docs.

• Scenarios end as Pass, Fixed, Skipped, or a Blocked state.
• Blocked (needs human verify) marks a check that depends on an outside service the branch changed and no local stand-in covers.
• Blocked (human decision) marks a fix too large to make autonomously.

What do you need before running it?

You need a local dev server the agent can start, such as bin/dev, rails server, or npm run dev. You need either a browser built into your agent or the agent-browser CLI installed. The skill runs scripts, so Python 3 is required to run scripts/packs-resolve.py. The repo needs Git, with its trunk branch available for diffing.

Compound Packs are optional. If your repo declares them in .compound-engineering/config.yaml, the skill uses them to get personas and UX criteria. Without them, personas are inferred or taken from your product docs, and the report says so.

• A local dev server the agent can start
• A built-in browser, or the agent-browser CLI
• Python 3
• Git, with the trunk branch available
• Optional: Compound Packs declared in .compound-engineering/config.yaml

What are the safety notes?

Read these before you run it. The skill runs scripts and uses the network. It commits fixes itself, so commits can land without review if you do not check them afterwards. It is scoped to the branch diff by design, so it does not test the whole app.

If you do not want the agent committing to a repo, do not use this skill there. Projects without a runnable local dev server will not work, because the browser run needs the app running.

• Runs scripts and uses the network.
• Commits fixes automatically, so review the commit history after a run.
• Screenshots and temporary files go to the OS temp folder, not your repo root.
• Only one browser is used for the whole run.