Eight engineers on a platform team each hand write their own prompt when they dispatch a Task subagent to explain a legacy module, and each chooses the subagent's tools ad hoc. Over one month the reports arrive in six different shapes, two subagents were given Bash and spent turns installing dependencies the task did not need, and the average time a reviewer spends reading a report has risen from four minutes to eleven. The team lead wants every module explanation to arrive in one shape with one tool set. Which change most effectively addresses this?
- AAsk each engineer to copy an agreed prompt template from the team wiki into the Task call, and review the pasted prompts during the weekly code review to catch any that have drifted.
- BDefine a single reusable AgentDefinition for the module explainer carrying its prompt, its description and its tool set, and have every engineer invoke that definition rather than composing a prompt. Correct
- CRecord the required report shape in the project CLAUDE.md so that each subagent picks the convention up from the repository, and leave the tool set to the engineer dispatching the task.
- DAdd three worked examples of a well formed module explanation to each engineer's prompt, so that the expected headings and level of detail are demonstrated before the subagent begins work.
Why A is wrong: A shared template does raise consistency and needs no engineering work, but it depends on eight people pasting the current version correctly every time, and a review held after the subagent has run cannot prevent the misconfigured tool set from being used.
Why B is correct: An AgentDefinition makes the prompt and the permitted tools one versioned artefact that every invocation resolves to, so the report shape and the tool set stop being per engineer choices and a change made once reaches all eight callers.
Why C is wrong: Repository level guidance is genuinely useful for conventions and travels with the code, but it is guidance rather than configuration and it leaves the tool set unaddressed, which is what allowed the two subagents to spend turns on dependency installation.
Why D is wrong: Examples do reliably tighten output shape when a model is unsure what is wanted, but they must still be pasted into eight separate prompts to take effect, so the drift they are meant to cure reappears through the same route it arrived by.