📚 Learning
Decode Graded Exams to Find and Fix Recurring Mistake Patterns
Upload photos of your graded exams to Claude's free tier and get a structured error-pattern analysis showing exactly which cognitive mistake types you repeat — no credit card, roughly 10–15 messages per 8-hour window.
Verified 2025-07-19
T0: Free, No Card
~10–15 msgs / 8hr
Image upload
6 error types
Concept Gap
Didn't understand the underlying idea
Calculation
Right concept, wrong math
Reading
Misread the question or own work
Recall
Couldn't retrieve a fact or formula
Application
Knew it, couldn't apply it here
Partial
Got partway, couldn't finish
📊 Message Budget Calculator
4 of ~12 messages used
~8 remaining
Step-by-Step Walkthrough
What Breaks
⚠ Known Limitations
- ~10–15 messages per 8-hour window — a single exam analysis uses 1–3 messages, so you can process roughly 3–5 exams per day. Hitting the cap mid-analysis means waiting for the window to reset.
- Photo quality matters — blurry images, faint grader marks, or cluttered annotations reduce accuracy. Photograph on a flat, well-lit surface with no shadows.
- Claude infers your answers from what's visible — if you erased and rewrote, or if the grader's marks are ambiguous, the classification may be wrong. Cross-check a few entries.
- The taxonomy is AI-interpreted, not ground-truth — Claude cannot read your mind. A "calculation error" might actually be a concept gap you haven't recognized. Use the taxonomy as a strong hypothesis, not a final diagnosis.
- No integration with Anki, Notion, or LMS platforms — output is plain text in the chat. Copy-paste into your study tools manually.
- Doesn't replace doing practice problems — the analysis tells you what to study, not how to study it. You still need to put in the reps.
- Limited to what's on the exam — if the test didn't cover certain concepts, those gaps won't surface. Don't assume mastery just because something didn't appear as an error.
- Free-tier model may change — Anthropic occasionally adjusts which model powers the free tier. Classification quality may vary if the underlying model shifts.
Three Things Worth Knowing
Why six error types instead of just "right" and "wrong"?
+
Cognitive research on metacognition shows that students who classify why they got something wrong improve faster than students who just re-study the topic. A calculation error needs drill practice; a concept gap needs re-learning from scratch; a reading error needs slowing down. If you treat all three the same way — "study more" — you waste time on the wrong remedy for the wrong disease.
Can't I just do this myself by reviewing my exams?
+
You can, and you should cross-check the AI's output. But most students don't — they look at the score, feel bad, and move on. Even when students do review, they tend to classify every error as "I didn't study enough" rather than distinguishing concept gaps from calculation errors from reading mistakes. The AI forces the distinction, which is where the diagnostic value lives.
What if I disagree with Claude's classification?
+
Tell it. Say "Question 4 wasn't a calculation error — I actually didn't understand the chain rule. Reclassify and update the summary." Claude will adjust. This is also a useful metacognitive exercise in its own right: arguing with the classification forces you to articulate what actually went wrong, which is the whole point of the workflow.
Open Claude, upload your last graded exam, and find out which mistake type is costing you the most points.
Open Claude →