use

Review and everyday use

How should I review, run work in parallel, and use wait time?

In one minute

Review is now the bottleneck, not writing code. Give it real time and priority.
Use AI to review faster and see more: understand, then review. Neither replaces the named approver.
Keep a short queue. Finish one task; two or three only when they are small and already waiting on someone else.
Fill waits with review, not with new tasks.

Why review comes first now

Before agents, producing the change was the slow part and review fit in a short daily slot. That split no longer holds. Agents produce diffs faster than anyone can responsibly merge them, and a green run is not a merge signal: in METR's study, roughly half of test-passing AI pull requests would not have been merged by their maintainers. See the verification gap.

The scarce work is now:

understanding what changed and why;
judging whether it is the right change, not only a green one;
reviewing other people's pull requests at the same increased volume.

Stageway can open a draft pull request. It never approves or merges one.

Checkpoints REV-3 REV-1

Review with AI assistance

Skill	Use it to
`review`	Review your own branch against the requirements and the actual diff. Findings are advisory; in a direct invocation it makes no edits unless you ask.
`understand`	Learn a change deeply enough to judge it. It teaches one concept at a time and checks your understanding before moving on.
`review-guide`	Get a reading order through a finished change: the key decision, the risky boundaries, the strongest tests, and reproducible checks.
`peer-pr-review`	Review someone else's pull request read-only, and prepare feedback that is not sent.
`peer-rfc-review`	Do the same for a design proposal.
`request-review`	Draft a short review request. It sends nothing without your explicit authorization.
`address-pr-comments`	As the author, retrieve every comment and thread, apply the fixes you select, and draft replies. Nothing posts without approval.
`pr-preflight`	Check a branch before it goes out: scope, evidence, credentials, and a PR body that does not claim testing that never happened.

The useful pattern is understand, then review. understand raises the quality of the human judgment. review and peer-pr-review widen what that judgment can see. Neither replaces the named approver.

Prefer a short queue

Worktrees isolate parallel checkouts. That is an isolation feature, not a throughput target. Stageway also admits only one runner per project, held by an operating-system lock, so concurrent runs can't fight over the same files.

Running many tasks at once produces:

a pile of pending draft pull requests;
merge conflicts against the base branch and against each other;
review debt that grows faster than any one task.

Rule of thumb

Finish one task at a time. Two or three are reasonable when they are small, independent, and already waiting on CI or a reviewer.

This is a limit on attention, not on tools. The runner can still work unattended; you just don't start the next task only because you can.

Checkpoints EXE-5

Where a task waits

specify
    → short wait while the agent researches
    → you answer questions and confirm the plan
implement-pipeline
    → long wait: tests, implementation, optional review, optional delivery
    → you review the result
    → wait for CI and peer review
    → address comments or fix CI
    → wait again as needed

The manual lane shortens the middle wait because implementation stays in your session. It does not remove the CI and peer-review waits, or the need to review the diff yourself.

What to do in the wait

Wait	Better use of the time	Worse use of the time
After specify, before you confirm	Read the plan. Check scope, checks, and exclusions.	Start a second specify on an unrelated task.
While the pipeline runs	Review an open pull request, yours or a colleague's. Run `understand` on a change you must approve soon.	Start another full pipeline.
After a draft pull request	Review this diff with `review-guide` and `understand`. Watch CI.	Open the next issue in a new worktree.
Blocked on comments or CI	Address the comments. Diagnose CI. Stay on the same change.	Start another task "so you are not idle".

Starting a new task during a wait is justified only when all three are true:

the current task is genuinely idle, waiting on someone else, with nothing you can act on;
you are still within the one-to-three task limit;
the new task is small enough not to create another long-lived draft pull request.

Otherwise, review.

Everyday patterns

One medium task, full pipeline. Specify, confirm, and start implement-pipeline. Spend the long wait reviewing another pull request. When the run completes, review this diff before asking a colleague. Start the next medium task only once this one is in CI or with a reviewer.

Two or three small follow-ups. Use the manual lane. Keep each change independently mergeable. Switch only when one is waiting on CI or a reviewer.

What not to do. Several worktrees, several draft pull requests, and a day spent resolving merge conflicts. The isolation works. The queue doesn't.