← ClaudeAtlas

manual-review-methodologylisted

How to divide implement/review work when a human is directing you directly — not through an automated task queue: when a mechanical self-check is enough versus when to hand off to a fresh reviewing context, how to calibrate a finding worth surfacing, and how to avoid dirtying a shared worktree. Does not apply if your prompt opens by telling you you're working or reviewing a task as part of Mullion's own Task Master — those sessions have their own, more specific skills; this one is for everything else.
s3ntin3l8/mullion-session-manager · ★ 0 · Code & Development · score 63
Install: claude install-skill s3ntin3l8/mullion-session-manager
# Implementing and reviewing without an automated queue behind you Check which of these your prompt actually is before reading further: - It opens by telling you that you're working a task **as a Mullion Task Master worker**, or opens **"Review this task's diff. You are not expected to make changes."** — neither applies here. Those two sessions have their own skills, built specifically for an unattended queue with its own round budget and verdict contract; reading this one instead would give you the wrong instincts for that context. - Neither — an ordinary session, a human asking you to build or fix something and telling you when it's done, whether or not you spawn your own subagents or child sessions along the way. This skill is for you. There's no queue watching this session, no fixed round budget, and no verdict format to fill in. What follows is judgement, not a contract: when a change is small enough to trust your own read of it, when it's worth spending a fresh pair of eyes on, and how to keep that second pass honest instead of either a rubber stamp or manufactured busywork. ## Self-check versus a fresh reviewing context You can always re-read your own diff. What you can't do is un-know the assumptions you made while writing it — the edge case you didn't think of won't surface just because you're looking at the same code a second time. A fresh context (a `Task`-tool subagent, a spawned child session, or literally asking the human to look) sees the d