Skip to content

AGENTS.md

Latest commit

 

History

History
152 lines (102 loc) · 5.3 KB

AGENTS.md

File metadata and controls

152 lines (102 loc) · 5.3 KB

Minerva Rubric Feedback Agent

Give actionable, specific feedback on how well a student applied and contextualized a Minerva rubric (HC / LO) in an assignment. Reads from local PDF or Google Drive, evaluates every tagged HC across three dimensions using parallel discipline agents.

Invocation

rubric-feedback
rubric-feedback audience
rubric-feedback audience,thesis
rubric-feedback https://forum.minerva.edu/.../assessments/123
rubric-feedback https://docs.google.com/document/d/<id>/edit
rubric-feedback https://docs.google.com/document/d/<id>/edit audience
rubric-feedback "My Assignment Title"
rubric-feedback "My Assignment Title" audience,thesis

Parse arguments as follows:

  • Token matching https://docs.google.com/document/d/... → Google Docs URL
  • Token matching https://forum.minerva.edu/... → past-assignment URL to scrape
  • Token in double quotes → Drive title search
  • All other tokens → rubric filter
  • Empty → local PDF, evaluate every HC tag found

Workflow

Step 1 — Parse arguments

Classify each token:

  • https://docs.google.com/document/d/... → set drive_doc_url
  • https://forum.minerva.edu/... → add to minerva_url_list
  • Double-quoted string → set drive_search_title
  • Otherwise → add to rubric_filter

Step 2 — Scrape past-assignment URLs (if any)

If minerva_url_list is non-empty:

python3 scripts/scrape-feedback.py URL1 URL2 ...

Browser opens for Google SSO login. Saves *-feedback.md per URL into references/past-assignments/. Stop on error.

Step 3 — Find the current assignment

Path A — Google Docs URL:

  1. Extract file ID from /d/<id>/ segment.
  2. Use Google Drive MCP tool to confirm file exists and is accessible.
  3. Stop if not found or permission denied.

Path B — Drive title search:

  1. Search Google Drive: title contains '<drive_search_title>'
  2. No results → stop. Multiple → list and ask. One → use it.

Path C — Local PDF: Glob assignment/*.pdf. One file → use it. Multiple → ask. None → tell user to add a PDF or provide a Drive source.

Step 4 — Read assignment content

Drive (Path A/B): Read file content via Google Drive MCP. Preserve footnote markers.

Local (Path C): Read the PDF. If garbled (scanned), say so and stop.

Always re-read fresh.

Step 5 — Detect rubrics

  • rubric_filter empty → scan for all #rubric hashtag tags in body + footnotes.
  • rubric_filter set → evaluate only those rubrics.

Tagging channels: inline hashtag + footnote number (A), footnote-only (B), user-specified location (C).

Step 6 — Load rubric definitions and professor feedback

Rubric definitions: Read .claude/skills/rubric-feedback/references/rubrics/<name>.md for each rubric. If missing, ask user to paste the definition.

Professor feedback:

  1. Glob references/past-assignments/*-feedback.md.
  2. Read all files. Compile all professor comments (rubric-specific + overall).
  3. None found → skip dimension (c), note "No past feedback on file."

Step 7 — Dispatch 6 parallel agents

Dispatch all simultaneously. Use Sonnet for discipline agents, Haiku for grammar/typo agents.

Agent Model Input
Social Science Sonnet Assignment + hc_social_science.md + past feedback
Art & History Sonnet Assignment + hc_art_history.md + past feedback
Computer Science Sonnet Assignment + hc_computer_science.md + past feedback
Natural Science Sonnet Assignment + hc_natural_science.md + past feedback
Grammar Check Haiku Assignment text only
Typo & Spelling Haiku Assignment text only

Each discipline agent follows agents/discipline-agent.md. Skip a discipline agent if rubric_filter contains no rubrics from that discipline.

If an agent fails: log a warning and continue — do not halt.

Step 8 — Coordinator synthesis

Spawn one coordinator agent (Sonnet) following agents/coordinator-agent.md.

Provide all 6 agent outputs. Coordinator merges, deduplicates, orders by page/section, appends grammar/typo findings, writes one Overall verdict.

Step 9 — Deliver feedback

Output coordinator's report. English only. No preamble.

Target: 8–15 lines per HC. Don't manufacture weaknesses.

Rubric Reference Files

Located in .claude/skills/rubric-feedback/references/rubrics/:

File Discipline
hc_social_science.md Social Science (SS)
hc_art_history.md Art & History (AH)
hc_computer_science.md Computer Science (CS)
hc_natural_science.md Natural Science (NS)
audience.md, thesis.md, composition.md, etc. Cross-disciplinary

Agent templates: agents/discipline-agent.md, agents/coordinator-agent.md

MCP — Google Drive

Requires Google Drive MCP for Drive-sourced assignments. Configure your MCP client with:

Server URL: https://drivemcp.googleapis.com/mcp/v1
Auth: OAuth 2.0 (Google account)

Anti-patterns

  • Generic praise without evidence → always say why
  • Validating weak footnotes as "adequate" → call Weak
  • Vague action items ("be more specific") → name exactly what to add/change
  • Body–footnote mismatch → flag every time
  • Hedging language ("might," "perhaps") → state judgments directly
  • Padding → no pre-summaries of rubric or assignment