exam-grader

name: exam-grader description: Grade a completed mock exam answer sheet against the question bank. Read-only — never modifies files. allowed-tools: [Read, Glob] argument-hint: " answers-string-like-1:A,2:C,3:B,..."

exam-grader skill

Takes a rendered mock-exam file and the user's answer string, produces a per-domain, per-scenario accuracy report.

The skill is allowed-tools: [Read, Glob] only — it cannot write or execute code. This intentionally limits its blast radius (pure scoring function) and demonstrates Domain 3.2's allowed-tools frontmatter restricting skill execution.

Inputs

• $1 — absolute or repo-relative path to the mock-exam markdown file (output of render_mock.py or /mock-exam).
• $2 — the user's answers as a comma-separated list of N:LETTER pairs, or one letter per line.

Procedure

1. Read the mock-exam file. Parse it into an ordered list of question IDs — each question block has a visible id: header or an HTML comment like <!-- id: q-cs-001 -->.

2. For each question id, read the corresponding canonical file under practice/questions/by-scenario/**/q-<id>.md. Extract correct, domains, scenario, difficulty.

3. Parse the user's answers string into a map { question_number: letter }.

4. For each question:

   • Mark correct/incorrect.
   • Accumulate per-domain counts (one question can count against multiple domains).
   • Accumulate per-scenario counts.
   • Accumulate per-difficulty counts.

5. Compute:

   • Raw score (correct / total).
   • Weighted-scaled estimate: (correct × 1000 / total). Output with a caveat that this is not the official scaled conversion.
   • Per-domain accuracy (correct / seen, per domain).
   • Per-scenario accuracy.
   • Per-difficulty accuracy.

6. Return a structured markdown report including:

   • Raw + scaled-estimate.
   • Pass/fail against the 720 threshold.
   • Per-domain and per-scenario tables.
   • A list of the question IDs answered incorrectly, so the user can re-read each.

Anti-patterns

• Do not recommend a specific pass threshold other than 720 (Anthropic's published minimum).
• Do not try to write to the study log — that's the caller's responsibility.
• Do not execute code or fetch network resources.