Start or continue agent work

Start an agent when you can name the job, the working directory, and the check that will tell you whether the work is useful.

Continue an agent when the next step is clear from the task, plan, session log, or diff. If you are guessing what the agent was doing, stop and rebuild the context first.

Before you start

Check four things:

1. Project — the repo is the one you mean to edit.
2. Branch or worktree — the agent has the right lane of work.
3. Task — the work has a durable record if someone else needs to review it.
4. Agent setup — the CLI works in a normal terminal.

A plain brew is fine for a small question. Use a task-linked brew when the work may create a diff, need review, or outlive the current terminal session.

Write the first prompt

A good first prompt gives the agent a job, a boundary, and a check.

Read the task and inspect the docs app. Update only the workflow docs that are still stubs. Keep the voice plain and run the docs lint command when done.

That prompt is not fancy. It is useful because it tells the agent what to do, what not to do, and how to prove the change.

Watch for the first risky moment

You do not need to stare at every line of output. Watch the moments where risk changes:

• The agent proposes a plan.
• It asks for permission to run a command.
• It starts editing files outside the expected area.
• A command fails.
• It says the work is done.

Interrupt early when the direction is wrong. A short correction before edits is cheaper than reviewing a wandering branch.

Continue from a saved or existing session

Before continuing an older brew, read the record:

1. Read the task description.
2. Read the latest plan if one exists.
3. Read the relevant session logs.
4. Inspect the current diff.
5. Check the active path and branch.

Then send a prompt that names the next action. Do not ask the agent to “continue” if there are several possible meanings.

Useful:

Continue from the latest session log. Fix the two review comments only, then rerun the docs checks listed in the task.

Less useful:

Keep going.

When to start fresh

Start a new brew instead of continuing when:

• The old session has too much stale context.
• The branch was rebased or heavily changed by someone else.
• You want a review-only pass with no edits.
• The next task is related but not the same task.

Fresh context is often safer than a long session carrying old assumptions.

Finish the turn cleanly

Ask for a short end note when the work matters:

Before you stop, summarize what changed, what checks ran, and what still needs review.

Then verify the result yourself. The agent's summary is a map, not proof.

Next: make a code change or read the agent model.