I'm a 9th-year Social Studies teacher. Every Sunday I write lesson plans. I've been doing it for 9 years.
At some point I had a folder with 246 files. Unit plans, daily lessons, worksheets, DBQ packets, sub plans. Nearly a decade of craft, structure, and pedagogical DNA sitting in a Google Drive folder nobody would ever read again.
I started wondering: what if an AI could read all of that and understand how I teach? Not just what I teach, but the specific way I open a class, structure a discussion, scaffold a concept, write a do-now?
So I built Claw-ED.
What it does
You point it at your lesson folder. It ingests everything: PDFs, DOCX, PPTX, plain text. It builds a "teaching fingerprint": your vocabulary, your structural patterns, your tone, your pedagogical approach.
Then when you ask it for a lesson on WWI or the Civil Rights Movement, it generates something that actually sounds like you wrote it. Not generic AI output. Not a curriculum company's voice. Yours.
Real output from one teacher's American Revolution unit:
"Alright, friends, as you settle in, I want you to take out your notebook
and answer this question: 'What does freedom mean to you? Is there ever
a time when following the rules is more important than being free?'
Take 5 minutes to jot down your honest thoughts."
The warm "friends," the specific structure, the invitational framing: all extracted from that teacher's existing materials. Not prompted. Learned.
How it works (technically)
Two phases:
Phase 1: Persona extraction
- Chunk and embed the teacher's documents
- Structured analysis: lesson structure patterns, vocabulary level, pedagogical markers, assessment approach, differentiation patterns
Phase 2: Guided generation
- Build a persona context string from the extracted profile
- Inject it into every generation prompt
- Output is voice-consistent because it was calibrated on that teacher's actual work
What's built
-
clawed chat- terminal chat interface -
clawed serve- FastAPI web dashboard -
clawed bot --token TOKEN- Telegram teacher bot -
clawed ingest <path>- learn from your lesson files -
clawed unit "Topic" -g 8 -s "Social Studies"- generate a unit plan -
clawed lesson "Topic"- generate a single lesson -
clawed standards list -g 8 -s math- browse state standards (all 50) -
clawed gap-analyze- find curriculum gaps vs. standards - IEP/504 differentiation engine
- Student chatbot (answers in teacher's voice, 24/7)
- PPTX/DOCX/PDF export
- MCP server (callable from any agent)
Privacy first
Files never leave your machine. Runs fully offline with Ollama. API keys in OS keychain. No telemetry.
Try it
pip install clawed
clawed demo # no API key needed
GitHub: https://github.com/SirhanMacx/Claw-ED#-getting-started
Would love feedback from teachers, ed tech folks, or developers who know what classrooms actually need.
Top comments (2)
This is a really smart approach to the "AI sounds generic" problem. The persona extraction phase — building a teaching fingerprint from vocabulary patterns and structural habits — is essentially what every domain-specific AI tool should be doing but most skip.
One thing I've been thinking about in the education space: the biggest bottleneck for independent tutors isn't lesson planning, it's discovery and payments. Teachers build incredible materials but have no marketplace layer to monetize 1-on-1 time outside of institutional settings. The next wave is probably tools like this paired with direct student-to-tutor matching.
How are you handling drift over time? If a teacher's style evolves (say they shift from lecture-heavy to Socratic method), does the fingerprint recalibrate automatically or does it need a fresh ingest?
Really cool approach with the voice fingerprint concept. Preserving a teacher's unique style is something edtech has largely ignored — most AI tools flatten everything into the same generic tone, which defeats the purpose when teaching is so personal.
The offline/privacy-first angle is smart too. Teachers shouldn't have to upload their entire lesson library to some cloud just to get AI assistance.
I think this idea has serious legs beyond classroom content generation. In 1-on-1 tutoring, matching communication styles between tutor and student is one of the biggest factors in outcomes. A tutor who explains things visually vs. one who uses analogies vs. one who's very Socratic — being able to capture and match those styles could transform how tutoring platforms connect people. We're exploring some of these personalization ideas over at tutorlinkup.com (crypto-friendly tutoring marketplace), and the "voice" problem is real.
Would love to see how Claw-ED handles different subject domains. Does the voice model adapt differently for, say, a math teacher vs. a humanities teacher?