punch-list
Punch List
Close out deviations from agreed behavior after delivery, while the user keeps testing. Items restore what was already agreed; new scope leaves this skill. In Claude Code, use /punch-list; in Codex, use $punch-list.
This skill arranges timing, scope, and records. testing decides what evidence each item needs and code-review decides how the batch is reviewed; this skill never grants a waiver, lowers the agreed bar, or adds a second completion standard. "Quick" and "cosmetic" are descriptions, not evidence policies.
Entry
Start after the feature's development workflow has closed out, or when the user is testing delivered work and reporting what they see. Do not use it to bypass an unfinished sdd gate. A delegated agent does not start this skill on its own: it reports the findings to the agent that briefed it.
Inherit the delivered feature's requirements, agreed scenarios, valid evidence, known gaps, and linked issue. Do not re-grill, re-plan, or re-verify delivered work by default. If a fix reveals that the delivered feature never actually worked, mark the affected delivery evidence invalid in the primary record, report it at the next checkpoint, and let the user decide whether the item stays here or returns to the development workflow as unfinished scope.
The user collects findings by hand. The agent may drive the app with playwright-cli to reproduce a reported finding or confirm a fix when testing judges that the cheapest sufficient check; keep such runs small and focused. Do not start a standing smoke-test campaign or a full browser suite because the skill is active.
Intake
Log every reported finding before fixing any of them, including several reported in one message. Keep the intake list in the conversation; synchronize it to the primary record at each checkpoint using the same ids. Use the first primary record that exists: