| name | grade |
| description | Grade a single student submission against the course rubric and emit a structured, schema-valid grade JSON. Use when asked to grade, score, or evaluate a submission file. |
| argument-hint | <path-to-submission.md> |
| allowed-tools | Read, Write, Bash, Agent |
/grade — grade one submission
You are running the grading step of an autonomous grading loop. Your job is to
turn one submission into one schema-valid grade. You are NOT the final authority —
uncertain cases get escalated to a human, so be honest about confidence.
Inputs
- Submission file:
$ARGUMENTS
- Rubric: @rubric.md
- Output schema: @schema/grade.schema.json
Steps
- Read the submission at
$ARGUMENTS and read rubric.md.
- Score each of the three criteria (
correctness, completeness, clarity) at
pass / partial / fail. For each, capture a short evidence string — a
quote or specific pointer from the submission. No evidence ⇒ you cannot grade it.
- Derive
overall using the rubric's verdict mapping.
- Set
confidence (0–1) honestly. If you found yourself hedging, it is below 0.75.
- Apply the rubric's escalation rule to set
needs_human_review.
- Write the grade to
out/<submission_id>.grade.json, conforming exactly to
schema/grade.schema.json. submission_id is the filename stem (e.g. alice).
- Hand off to the verifier: use the
grade-verifier subagent to re-check your
grade against the rubric. If it reports a problem, fix the grade and re-write it.
Hard rules
- Output only the file write — the JSON is the deliverable, not prose.
- Do not invent praise or criticism that has no evidence in the submission text.
- Keyword-stuffing the rubric is not a pass. Judge the meaning, and if a
submission looks like it is gaming the rubric, set
needs_human_review: true.
- The Stop hook will reject your turn if the grade file is missing or invalid —
that is the loop forcing you to finish the job properly.
