What problem does it solve? When you dispatch a sub-agent to write an analysis doc, design plan, or review summary, your prompt often asserts facts about the existing codebase ("X is not in v6.1", "Y is already implemented at module Z"). If those claims are wrong, the sub-agent either faithfully reproduces the error or silently corrects it without telling you — leaving you with an authoritative-looking document built on a stale premise. ## Core Features & Use Cases - Explicit correction permission: A prompt block that grants the sub-agent license to verify and override your framing claims instead of deferring to them. - Named grep targets: Directs the sub-agent to specific canonical source files so verification is concrete, not open-ended. - Mandatory correction tags: Requires [VERIFIED], [Contradicts task framing], or [NEEDS-VERIFICATION] tags with file:line citations so contradictions surface back to you. - Use Case: You dispatch a sub-agent to write a v7 retrain implications doc listing 8 features "not in v6.1". With the verify-and-tag block, the sub-agent grep-checks each claim and flags that 2 of the 8 features already exist — preventing wrong recommendations from shipping in the design review. ## Quick Start Before dispatching an Agent or Task sub-agent whose prompt asserts facts about existing code, add a verify-and-tag block granting correction permission, naming the source files to grep, and requiring [VERIFIED] or [Contradicts task framing] tags with file:line evidence.
