cheatsheet-creator

Cheatsheet Creator

Build a focused, exam-ready markdown cheatsheet from a student's course materials.

Why this skill exists

Students often arrive with a pile of lecture slides, problem sets, and past exams and want them distilled into something they can actually study from or bring into an exam. Two things make this hard:

1. Memory. Course materials are big. Twelve lectures × 60 slides + eight homeworks + three past exams will blow past the working context if loaded all at once.
2. Signal vs. noise. Lectures cover everything; exams reveal what actually matters. The sharpest cheatsheets are anchored in what's been tested.

This skill addresses both: parse one file at a time, persist intermediate summaries to disk, then use past exams and homework as a "what to emphasize" filter applied at the end. The order matters — do not try to do this all at once.

Workflow

Step 0: Inventory, categorize, and confirm scope

Before parsing anything, list every input file and confirm with the user which are:

• Lectures — slides, notes, scanned handwritten notes, transcripts
• Homework — problem sets, ideally with solutions
• Past exams — midterms, finals, practice exams, ideally with solutions

Suggest one of these two organizing conventions so categorization is unambiguous (and so this skill can be re-run cleanly later):

• Folder structure: separate lectures/, homework/, exams/ directories.
• Filename prefixes: lecture_03_continuity.pdf, hw_05.pdf, exam_midterm_2023.pdf — any consistent prefix works.

If the user's files already follow one of these, proceed silently. If they don't, ask the user to confirm classifications and gently suggest organizing this way next time. Don't reshuffle their files for them.

Misclassifying matters: a homework problem summarized as "lecture content" pollutes the master concept list with one student's wrong-turn intermediate work.

Also confirm with the user:

• Target size. Always ask — the user told us they want this configurable each time. "One page front-and-back," "three pages," "exhaustive study guide," etc. This drives how aggressively to compress in the final synthesis step. Note that worked examples (Step 4) count toward this size budget.
• Cheatsheet rules, if any. Some exams allow only handwritten cheatsheets, some allow typed. The user might also have their own constraints like "formulas only, no worked examples."
• Subject/course name, so the final file can be named sensibly and the writing tuned (heavy LaTeX for a real analysis class vs. heavy code blocks for a data structures class).
• Save intermediates? Whether to include the per-lecture summaries, master concept list, exam frequency tally, and worked-examples scratch in the final output folder (under cheatsheet/intermediate/). Default to yes — they're useful while studying (you can see why a concept made it onto the sheet, audit the source attributions, and re-run the synthesis at a different size budget without reparsing). Skip them only if the user explicitly says they don't want them.
• Self-test rigor (optional — default: no holdout). The finished sheet is self-tested in Step 5b regardless, and that test uses a no-sheet control arm to measure how much the sheet itself contributes (vs. what the model already knows), so it doesn't need a held-out exam to be honest. By default, therefore, build the sheet from all of the student's exams (past exams are scarce — don't burn one) and let Step 5b run in-sample. Only if the user has ≥2 exams and explicitly wants a stricter out-of-sample estimate ("how will this do on the real final?"), offer to reserve one exam as a held-out exam excluded from Step 3. That trades a little of the sheet's emphasis signal for an unbiased generalization number. If reserved, pick a representative exam that has solutions (the grader needs them) and state in your plan which one is held out.