Education Review Desk
Education Review Desk は、intake 以外の強い例です。ユーザーはすでに artifact を持っていて、agent の仕事は承認済みの rubric や標準に沿って review し、役立つ feedback と revision plan を返すことです。

これはシンプルですが根拠のある review の形です。承認済みの rubric を Attachments に置き、draft を Live Test にアップロードして、agent に feedback と revision plan を返させます。
このパターンが向いている場面
- ユーザーが draft、case write-up、assignment をアップロードまたは貼り付ける
- チーム側に承認済みの rubric や review standard がある
- 出力は feedback、revision guidance、または human reviewer への escalation であるべき
Agent がすること
- 何を review しているのか、どの rubric を使うべきかを確認する
- 承認済みの criteria に沿って内容を評価する
- 強み、抜け、最優先で直すべき点を示す
- 範囲外なら revision plan を示すか、human reviewer に escalate する
承認済みの outcome
Feedback SummaryRevision PlanHuman Reviewer Escalation
してはいけないこと
- 承認されていない rubric や grading standard を作り出す
- artifact から裏づけできない問題を断定する
- judgment を説明せずに全体を書き換える
- human reviewer が必要な場面で最終的な academic / policy decision を出す
相性のよいツールとデータ
- Attachments: 小さな rubric、checklist、exemplar を常時参照させる時
- Knowledge Base: rubric library、policy note、curriculum guidance が大きい時
- Request Form: 後続フローで student や reviewer から structured follow-up が必要な時だけ使う
- HTTP Request: review result を LMS や内部 system に送る必要が出た時に追加する
会話例
User: climate policy essay を見てもらえますか? conclusion が弱い気がします。
Agent: はい。Ms. Chen の persuasive writing rubric に沿って review します。thesis clarity、evidence quality、counterarguments、conclusion impact を中心に見ます。
User: その rubric で合っています。
Agent: evidence section は強く、支えも十分です。ただ、conclusion は前半の繰り返しが多めです。いちばん大きい改善点は "so what?" をもっと直接答えることです。短い revision plan をまとめます。1) 主張を 1 文で言い直す、2) 現実の影響を明示する、3) 読者に支持してほしい policy action で締める、です。
このパターンが重要な理由
- Codeer が単なる first-touch intake ではなく、expert judgment を扱えることが伝わります。
- 教育、QA、compliance review、社内の expert feedback loop に向いています。
- 承認済みの基準と、境界を守った判断の見え方がより明確になります。