Back to catalog / Agent workflows

Grade a research brief with an independent rubric

Have a writer produce evidence, then let an independent grader check coverage and source quality.

Goal

Goal: Have a writer produce evidence, then let an independent grader check coverage and source quality.

Specify artifact path, format, and measurable acceptance criteria in the task; list each criterion separately in the rubric. [outcome]
Submit description and rubric through define_outcome, set bounded max_iterations, and require reviewable evidence in grader feedback. [outcome]
Read the terminal state (satisfied, max_iterations_reached, failed, or interrupted) and reconstruct the artifact from its final version. [outcome]

Check that each pass cites artifact/source evidence; each unmet criterion should trigger a targeted revision.

If writer and rubric disagree about the artifact path, the run can fail; keep task, rubric, and output paths aligned.

Python 3.11+, Anthropic API key, Managed Agents beta API, and authorized sources when research is required.

Make the rubric falsifiable; model grading is not financial, compliance, or legal approval.

An independent grader can still err; the cookbook requires evidence-oriented criteria and bounded iterations.

These values describe the upstream repository, not popularity or likes for an individual skill or playbook.

No metrics snapshot is available yet.

License: MIT