Canvas Homework

Canvas homework workflow ("Chip")

This is the harness. It connects four things: the canvas MCP server (lists and reads assignments), the scripts/canvas.mjs helper (downloads files the MCP can't), the built-in docx and pdf skills / Read tool (read the files), and the writing-voice skill (draft in Max's voice). Follow the steps in order.

Tools you will use

From the canvas MCP server (this is its complete toolset):

• canvas_list_active_courses — fast list of your current courses (dashboard API). Prefer this for "what am I taking right now."
• list_courses — all courses; use flags for completed/all. Use when you need a course that isn't on the active dashboard.
• search_assignments — search assignment titles/descriptions; filter by due-date range, course, and completion/submission status. Use this to find what's not done.
• get_assignment — full details for one assignment; ask for markdown output so the description and any embedded file links come through cleanly.

The canvas MCP cannot download attachments. That gap is filled by:

• node scripts/canvas.mjs assignment-files <courseId> <assignmentId> — start here. Finds every file linked from the assignment description, its attachments, and any file sitting in the same module. Dedupes and tells you where each one came from.
• node scripts/canvas.mjs files <courseId> [nameFilter] — list downloadable files in a course, optionally filtered by a substring of the filename. Follows pagination, so it sees past the first 100 files.
• node scripts/canvas.mjs download <fileId|url> [outPath] — download one file into output/ (resolves Canvas's signed download URL using the API token).

Reading files (all built in — do not install an MCP for these):

• .docx → the docx skill (extract full text; never edit the source file).
• .pdf → the Read tool (use the pages parameter) or the pdf skill.

Checking a finished draft:

• node scripts/lint-draft.mjs <path> — mechanically enforces the writing-voice hard guardrails (no em dashes, no markdown). Exits non-zero and prints a line:col for every violation. This is not optional; see Branch B step 4.

Step 1 — List outstanding assignments

1. Call canvas_list_active_courses to get your current courses (and their IDs — you'll need IDs later). Fall back to list_courses if you need a non-active course.
2. Call search_assignments filtered to not submitted / not completed, ideally within an upcoming due-date window (e.g. now → +30 days). If the filter options don't cleanly express "incomplete," pull the candidates and filter by submission status / due date yourself.
3. Show Max a clean list: course → assignment title → due date → points. Sort by soonest due. If he just asked "what do I have left," stop here and let him pick one. Don't auto-start drafting everything.

Step 2 — Pull the chosen assignment

Call get_assignment (markdown) for the assignment Max picks. Read the description in full. Note: the deliverable type (essay, discussion post, worksheet, short answer), the length/format requirements, the due date, and any rubric text.

Step 3 — Find a downloadable file (usually a DOCX)

Look for an attached file, cheapest path first:

1. Ask directly — node scripts/canvas.mjs assignment-files <courseId> <assignmentId>. This covers the two cases below in one call, including files that are only reachable through the assignment's module rather than the course Files tab. Usually this is the whole step.
2. In the description from get_assignment — scan for links to Canvas files (URLs containing /files/<id> or instructure_file_link anchors). Grab the file id or URL.
3. In the course files — run node scripts/canvas.mjs files <courseId> <nameFilter> and look for a file whose name matches the assignment (e.g. files 1234 essay). Drop the filter to see everything.

If you find one, download it:

node scripts/canvas.mjs download <fileId|url> output/<sensible-name.ext>

Then decide which branch you're in based on what the file actually is.

Step 4 — Branch on what you found

Branch A — a DOCX exists AND it contains the assignment

This is the "just show it to me" path.

1. Read the entire DOCX with the docx skill (extract text).
2. Do NOT edit the file. Leave it exactly as downloaded.
3. Print the full assignment text to the console (into the chat), verbatim and complete — not a summary. If it's long, print all of it anyway; that's the point.
4. After printing, offer (one line): "Want me to draft a response in your voice?" — if yes, continue into Branch B using this DOCX's contents as the prompt.

"Contains the assignment" means the DOCX is the actual prompt/instructions or a worksheet to read. If the DOCX is empty, or is just a rubric/cover sheet and the real task lives in the Canvas description, treat it as Branch B instead.

Branch B — no DOCX, the DOCX isn't the assignment, or Max asked for a draft

This is the "do it in my voice" path.

1. Figure out the task. From the get_assignment description, determine exactly what has to be produced: the deliverable type, the prompt/question, required length, format, citation style, and any rubric criteria. State it back in one or two sentences so Max can confirm you understood the assignment.
2. Read the context files. If the assignment references a reading or attaches a PDF (or any other format), download it via canvas.mjs download and read it with the Read tool (pages) or pdf skill. Pull out the facts, quotes, and specifics you'll need so the draft is grounded in the actual source material.
3. Draft in Max's voice. Load the writing-voice skill (.claude/skills/writing-voice/SKILL.md) and write the response following it: stance-first, myth-busting, concrete numbers/examples, his punctuation and idiom habits, register matched to the assignment (academic vs. business/opinion). Hit the required length and rubric points. Obey the writing-voice hard guardrails: the draft must contain no em dashes and no markdown formatting (plain prose only). Scan the finished draft for "—", "*", "_", "#", and backticks before printing.
4. Save, lint, then print — in that order:
   • Save it to output/<course>-<assignment-slug>.md.
   • Run the guardrail linter on it:

     node scripts/lint-draft.mjs output/<course>-<assignment-slug>.md
     

     It prints a line:col for every em dash and every markdown character, and exits non-zero if it finds any. Fix each one, save, and re-run until it prints Clean. Do not print the draft to Max until the linter passes. Your own scan in step 3 is the first pass; this is the backstop, and it is the one that actually decides.
   • Print the full draft to the console (into the chat).
5. Tell Max where it saved, the word count, and which rubric requirements it covers.

Step 5 — Hand-off note

End every Branch B draft with a short, honest reminder (one line, not a lecture): "This is a first draft in your voice. Read it, make it yours, and check your course's policy on AI assistance before submitting." This is Max's own writing style applied to his own assignments; the point is to beat the blank page and study, not to submit unread.

Guardrails & tips

• Never edit a downloaded assignment file. Read it; don't write back to it. Drafts go in output/, originals stay untouched.
• Print, don't just summarize, in Branch A. Max wants the whole thing surfaced.
• If the canvas MCP isn't connected, say so and fall back to node scripts/canvas.mjs courses / files to confirm credentials work, then point Max at README.md setup.
• If credentials are missing, canvas.mjs prints exactly what to set — relay that.
• Course IDs from list_courses are what canvas.mjs files <courseId> needs; keep track of them as you go.