Your first brew

Start with one small, reviewable change. The goal of your first brew is not speed. It is learning where the agent worked, what it changed, and how you verify it.

Pick a safe task

Good first tasks are narrow:

• Explain the repo structure.
• Update one docs page.
• Fix a typo or small UI copy issue.
• Add a test around an obvious bug.
• Review a diff without editing files.

Avoid broad prompts like "improve the app" or "refactor the frontend". Those prompts make bad review work because there is no clear finish line.

Start from the right place

1. Open the project in Coldtea.
2. Check the active branch and project path.
3. Open or create a task if this work should be reviewed later.
4. Choose the agent you want to use.
5. Start the brew.

Coldtea opens a terminal pane for the agent. The agent still owns its own prompt UI, model choices, slash commands, and permission prompts.

Write the first prompt

Give the agent a job, a boundary, and a check.

Read the docs app structure and update one getting-started page with practical setup steps. Keep the change small. Run the docs lint command when you are done.

That is better than:

Make the docs better.

The first prompt gives the agent a path to done. The second asks for taste, scope, and verification all at once.

Watch enough of the terminal

You do not need to stare at every token. You do need to catch the moments that change risk:

• The agent proposes a plan.
• It asks for permission to run a command.
• It starts editing files outside the task.
• A command fails.
• It claims the work is complete.

If something looks off, interrupt early. A small correction before edits is cheaper than reviewing a branch that wandered.

Review the result

Before you call the brew done:

1. Read the changed files.
2. Check the diff against the task.
3. Read the plan or session log if the brew is task-linked.
4. Run the relevant checks yourself or inspect the agent's check output.
5. Decide whether to continue, merge, or discard the work.

Trust the workflow. Verify the code.

A useful first habit

Ask the agent to leave a short session note when the task matters:

Before you stop, summarize what changed, what checks you ran, and anything I still need to review.

That note is not a substitute for review. It is a map for the next human who opens the task.

Next: review agent work or run parallel agents when the work can be split safely.