Why it matters
The agent that did the work is the worst one to judge it. It remembers what it meant to do, not what it actually did.
Case had a separate reviewer and closer that checked the implementer’s work. Riker requires an independent reviewer before a job counts as done. A worker can never verify itself.
One warning from Riker: two models agreeing is not proof. A reviewer is good at catching things you didn’t mean. Your checks are still what prove the work is correct.
Exercise: review cold
- Open a new Claude Code or Codex session, so it starts with none of the first session’s history.
- Give it only your brief and the change (the diff, the branch, or the files).
- Ask:
Here is the brief and the change. Don't fix anything.
List every way the change fails the brief, misses a "done when" line,
or changes something out of scope. Quote the exact lines.
Check yourself: for every finding, decide whether it’s real. Note which ones the first session’s summary didn’t mention.