Homework Grader Skill

Ask the user for the following if not already provided:

• Solution key: a PDF (typed) or image. This is the ground truth.
• Student submissions: one or more images per student, grouped by student. Ask the user how the students are identified (file name convention, folder, or manually labelled).
• Point values: Does each question have equal weight, or does the user want to specify per-question points? Default: equal weight, total = 100.
• Partial credit policy: Default policy is described in references/matching-and-grading.md. Ask only if the user wants to override.

If the user uploads files directly in the chat, accept them. If files are on disk, they will be at /mnt/user-data/uploads/.

Step 2 — Parse the Solution Key

Follow the detailed parsing guidance in references/parsing-strategy.md, Section 1.

Quick summary:

• If PDF: extract text directly using pdfplumber (preferred) or pymupdf. Fall back to vision if extraction yields garbled text (scanned PDF).
• If image: send directly to Claude vision with the structured extraction prompt in references/parsing-strategy.md.
• Output: a numbered list of questions with their canonical answers, e.g.:

Q1: Define P-wave velocity. Answer: Speed of compressional waves through a medium, ~6 km/s in continental crust.
Q2: Calculate depth given two-way travel time of 1.2s and v=3000 m/s. Answer: depth = (v × t) / 2 = 1800 m
...

Store this as solution_key: list[dict] with fields q_num, q_text, answer_text, points.

Step 3 — Parse Student Submissions

For each student, process their image set in order. Follow references/parsing-strategy.md, Section 2.

Key rules:

• Treat multiple images as ordered pages of a single document. Sort by filename if numeric suffix is present (e.g., student1_p1.jpg, student1_p2.jpg); otherwise ask the user for page order.
• Use Claude's vision to extract all handwritten content from each image.
• Concatenate extracted text across pages before attempting question detection.
• Extract student identifier from filename or ask the user.

Output per student: raw_text: str, detected_answers: list[dict] with fields detected_q_label (what the student wrote, e.g. "Q3", "3.", "iii", or None if no label), answer_text, page_number.

Step 4 — Match Student Answers to Solution Key

Follow references/matching-and-grading.md, Section 1 for full logic.

Two-pass matching strategy:

Pass 1 — Auto-detect: Try to map detected_q_label to q_num in the solution key.

• Normalise labels: "Q3", "3.", "3)", "iii", "(3)" all → 3
• If all detected labels map cleanly and cover ≥ 70% of solution key questions → use auto-detect mapping

Pass 2 — Content anchor fallback: If auto-detect fails (missing labels, ambiguous numbering, <70% coverage):

• Use semantic similarity: compare answer_text content to answer_text in solution key
• For math/equations: check if numerical values, variables, and formula structure match
• For geoscience: check for domain-specific keywords (velocity, impedance, reflection coefficient, etc.)
• Assign best content-match if similarity score > threshold (see references/matching-and-grading.md)

Flag the following as mismatches:

• MISSING: A solution key question has no matched student answer
• UNMATCHED: A student answer has no plausible match in solution key
• AMBIGUOUS: Two or more student answers could match the same solution key question
• REORDERED: Answer matched via content (not label) — note the original label vs. matched label

Step 5 — Grade Each Matched Answer

Follow references/matching-and-grading.md, Section 2 for rubric details.

Score each matched answer on a 0–4 scale (then scale to point value):

For math/equations:

• Check formula structure first, then numerical result
• Award partial credit for correct setup with arithmetic error

For geoscience/written answers:

• Check for required key concepts (listed in solution key)
• Each missing key concept deducts proportionally

MISSING questions → score 0 automatically.
UNMATCHED answers → score 0, flag for instructor review.

Step 6 — Build the Output Spreadsheet

Read /mnt/skills/public/xlsx/SKILL.md before writing the spreadsheet.

Follow the schema in references/spreadsheet-schema.md exactly.

Sheet 1: Grade Summary

• One row per student
• Columns: Student ID, Q1 score, Q2 score, … Qn score, Total Score, Total %, Flags (count of mismatches)

Sheet 2: Mismatch Flags

• One row per flagged item
• Columns: Student ID, Flag Type (MISSING / UNMATCHED / AMBIGUOUS / REORDERED), Solution Key Q#, Student Label, Notes

Sheet 3: Parsed Answers (Detail)

• One row per student × question
• Columns: Student ID, Q#, Student's Answer (transcribed), Expected Answer, Score (0–4), Points Awarded, Max Points, Grader Notes

Apply conditional formatting:

• Red fill for score = 0 in Grade Summary
• Yellow fill for REORDERED or AMBIGUOUS flags
• Orange fill for MISSING flags

Save to /mnt/user-data/outputs/homework_grades.xlsx and present to user.

Step 7 — Present Results

After writing the file, give the user a brief summary:

• Total students graded
• Class average score (%)
• Number of students with at least one mismatch flag
• Most commonly missed question (if any)

Call present_files with the xlsx path.

Error Handling

• Image too blurry to parse: Note in Parsed Answers sheet as ILLEGIBLE, score 0, flag for manual review.
• Student wrote in a language other than expected: Flag as LANGUAGE_MISMATCH, attempt translation, note uncertainty.
• Solution key has sub-questions (1a, 1b): Treat each sub-question as an independent question row.
• Duplicate student IDs: Warn the user and append _dup to the second occurrence.