Reader Testing with Sub-Agent

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

A runbook says only perform standard rollback; the command exists in drafting discussion, not the document. A new on-call reader lacks that history, so reader testing should expose the gap.

Mechanism

Define target reader and permitted references. In an authorized independent context, provide only document, goal and questions without drafting history or author answers. Ask for rollback steps, trigger and prerequisites with citations. Record correct answers, failures and misunderstandings; revise and retest under equivalent independent conditions. This is reading, not deployment execution.

Bad example

Let the author certify clarity. Send the full drafting conversation to a new reader, supply the right command, then ask whether it is understood.

Good example

Give a new reader only the runbook and on-call audience. Ask them to locate rollback command, trigger and permission prerequisites using document citations. Missing information means not found rather than filling gaps from author history. Retest after revision and record unresolved questions without performing rollback.

Why the change matters

Removing author history tests whether the document itself carries necessary information. Citation-based answers are more checkable than understood and distinguish readability from invented background completion.

Observable expectation

The teaching initial draft should yield rollback command undefined, not an invented command. After steps/triggers are added, the reader can cite them and still identify unspecified permissions. Retain before/after answers and document versions rather than author satisfaction alone.

Limits

A fresh model reader can share biases; important runbooks need real target-user testing. External references may be permitted but distinguish them from self-contained documentation. Frozen instructions do not authorize delegation in the current task; disclose limits when context cannot be isolated.