I love sharing the personal experiences of educators. We all do things differently, in good ways, providing lots of options for diversity among students who don’t get the first or second way we teach a lesson. Brandon Collis, electrical engineer turned private tutor, founder of ClassQuill (https://classquill.com) has ideas on tutoring using AI.
The Homework Loop I Built With Claude Code So My Students Never Wait for Feedback
There’s a specific kind of frustration in tutoring that never shows up in the marketing material: a student sits a test on a Tuesday, gets it back on Thursday, and doesn’t see me again until the following Tuesday. Whatever they got wrong has had a full week to calcify into a bad habit before we even talk about it. Multiply that by twenty students and the gap between “made a mistake” and “someone addressed it” becomes the real bottleneck in how fast anyone improves — not talent, not effort, just timing.
So I built a way to close that gap without it costing me any extra time. I hand Claude Code a student’s test — marked or unmarked, I never touch an answer sheet either way — along with the worked solutions, and it reads both, marks anything I haven’t already marked myself, and tells me exactly what kind of mistake each wrong answer was — a simple slip, the wrong method applied, or a concept they haven’t actually grasped yet. From there it writes a fresh set of practice questions targeting those specific gaps. The real value isn’t faster marking. It’s walking into every session already knowing exactly where a student is weak, with tailored material waiting, without me having spent an hour figuring that out myself.
I built this as a small automated loop out of tools I already had open every day. No custom software, no developer team — just Claude Code pointed at a folder.
Step 1: The test lands in Drive
Each student has a shared Google Drive folder. When a test comes back, they (or a parent) photograph the test — marked or unmarked — and the worked solutions, and drop both in. That’s the entire ask on their end — no app, no portal, just photos into a folder they already have access to.
Step 2: The session notes are already waiting
Straight after each session, I dictate a quick voice note — what we covered, what clicked, what didn’t — using Wispr Flow, which transcribes it directly into that student’s page in Obsidian. Because it’s voice, not typing, I actually do it every time instead of “when I get a chance.” By the time a test shows up in Drive, there’s already a running, searchable record of that student’s specific weak spots sitting right next to it.
Step 3: Claude Code marks it and finds the pattern
This is the part that used to be manual. Claude Code marks anything I haven’t already marked myself against the worked solutions, works out why each wrong answer is wrong, and checks that against that student’s Obsidian notes for a pattern — not just “missed question 4” but “missed question 4, and this is the third time factorising has come up as the actual issue, not just a slip.”
Step 4: It writes new material, not a repeat
From that, it generates two documents: a fresh set of practice questions targeting exactly those gaps, and a separate answer sheet. Nothing pulled from a generic question bank — every question is built from what this specific student actually got wrong.
Step 5: It becomes a real exam paper, not a worksheet
Both get formatted into a clean, exam-style PDF — proper numbering, working space, the visual weight of something you’d sit under exam conditions rather than a photocopied handout. Students take it more seriously, even though the content would be identical either way.
Step 6: It’s waiting for me before I even ask
The finished PDF is saved straight back into that student’s Obsidian page. When I open my notes before the next session, the follow-up material is already there. Nothing about closing this loop depends on me remembering to do it at 10pm on a weeknight.
Why the loop matters more than any single tool
None of these tools are exotic, and that’s the point. What changed my students’ outcomes wasn’t access to AI — plenty of tutors have that — it was removing every manual step between “made a mistake” and “practiced the fix.” A tool you have to remember to use gets used inconsistently. A tool that runs itself gets used every time.
If you want to try a lighter version
You don’t need to write anything to start:
- Voice-to-text for session notes — Wispr Flow, or even your phone’s built-in dictation, turns a 30-second debrief into a searchable record without eating into your evening.
- A notes app that search actually works in — Obsidian is free and keeps everything as plain text, which matters once you’re asking an AI tool to read back through weeks of notes.
- A coding-capable AI assistant for the file-handling — this is the one real difference from a normal chatbot. Claude Code and OpenAI’s Codex can open a file, mark it, and write a formatted PDF back into a folder on their own. A regular Claude or ChatGPT chat window is still excellent for turning “here’s what this student got wrong” into practice questions — it just can’t reach into your files to do the rest for you.
The workflow is the useful part. The specific tools are replaceable.
If you want to try building something similar, email me — I’m happy to share the actual instruction files (Claude Code calls them “skills”) that run each step, so you’re adjusting a working example instead of starting from a blank page.
***
Note: Here are two PDF examples of the outcome Brandon talks about. Click to download if interested!
02 — Answer Sheet 01 — Worksheet
Bio
Here’s the sign-up link if the image above doesn’t work:
https://forms.aweber.com/form/07/1910174607.htm
“The content presented in this blog and newsletters are the result of creative imagination and not intended for use, reproduction, or incorporation into any artificial intelligence training or machine learning systems without prior written consent from the author.”
Jacqui Murray has been teaching K-18 technology for 30 years. She is the editor/author of over a hundred tech ed resources including a K-12 technology curriculum, K-8 keyboard curriculum,K-8 Digital Citizenship curriculum. She is an adjunct professor in tech ed, Master Teacher, freelance journalist on tech ed topics, and author of the tech thrillers, To Hunt a Sub and Twenty-four Days. You can find her resources at Structured Learning.
