"Who has tried using ChatGPT or Claude for lesson planning
and got something that looked great...
...but was useless in your actual classroom?"
"LLMs predict the next token based on all previous tokens.
They don't plan, don't remember, and default to agreeable patterns."
"LLMs predict the next token based on all previous tokens"
What does this actually mean?
Let's see it in action.
Each token you write shapes what comes next.
Early words have cascading influence —
like setting GPS before driving.
Vague start ──────────────────────► Generic output
Specific start ──────────────────────► Specific output
(Front-load your specifics. Structure your start.)
"LLMs don't plan, don't remember"
The chat interface creates an ILLUSION of memory.
Here's what actually happens.
⚠ YOUR ORIGINAL "UNDER 5 POINTS" CONSTRAINT?
Still there... buried 650 tokens ago.
The model isn't "forgetting" your instructions.
Your constraints are still there — just buried under hundreds of newer tokens.
Long conversation ───────► Constraints diluted
Fresh chat + context ───────► Constraints intact
(Put critical requirements at the START. Repeat them in long conversations.)
"LLMs default to agreeable patterns"
Models are trained on human feedback. Humans prefer agreement.
So the model learns to say "Yes, great idea!"
Even when it's not.
AI: "This is an excellent hands-on project! It encourages curiosity and scientific thinking.
You could enhance it by adding graphing..."
Pure validation. No actual problems found.
AI: "Three failure points:
1. Students without 'controlled variable' concept will compare plants incorrectly.
2. 2-week timeline means absent students lose data.
3. 'Present findings' is vague — students won't know what to include."
Same project. One gives validation. One gives insight.
"Is this good?" ─────────────► triggers agreement patterns
"How will this fail?" ───────► triggers analysis patterns
The model CAN critique — but you have to ask explicitly.
(Ask for failure modes. Request critique. Override the default agreeableness.)
WHAT YOU PROBABLY GOT:
WHAT YOU PROBABLY GOT:
Here's a real student error:
Student (Class 8 Science):
"Water evaporates faster on hot days because heat makes water lighter, so it floats up into the air."
Try this prompt:
The "Statistical Leap" vs "Stepping Stones"
Long leap = High risk of hallucination
Short steps = High accuracy
WHAT YOU PROBABLY GOT:
"The student has a misconception about evaporation. You should explain that heat increases kinetic energy of water molecules, allowing them to escape. Try a demo with a hot plate..."
WHAT YOU PROBABLY GOT:
Diagnosis: "Student conflates 'lighter' (mass) with 'more energetic' (kinetic energy)."
Micro-task (90 seconds): "Put a drop of water on your hand. Blow on it. Does it get lighter? No. But it evaporates. What made it evaporate?"
Verification: "Can cold water evaporate? Why?"
Think of a lesson, project, or assessment you've designed.
Try this prompt:
Notice: Does the AI genuinely critique it?
Or does it validate and then add suggestions?
"This is an excellent project! It encourages hands-on learning and scientific inquiry.
You could enhance it by:
These additions would make an already strong project even better!"
| Student A | Fails at step 3 — doesn't know 'controlled variable' concept. Assumption: Taught in Cls 6 Fix: Add 5-min review |
| Student B | Disengages after Day 3 — no challenge Assumption: Observation alone is engaging Fix: Add hypothesis task |
| Student C | Confused by 'present findings' — what goes in a presentation? Assumption: They know Fix: Give template |
"LLMs predict the next token based on all previous tokens.
They don't plan, don't remember, and default to agreeable patterns."
| MECHANISM | SOLUTION |
|---|---|
| "Predict next token based on all previous tokens" | CONTEXT ANCHORING Front-load your specifics. |
| "Don't plan, don't remember" | CHAIN-OF-THOUGHT Force intermediate steps. |
| "Default to agreeable patterns" | ADVERSARIAL FRAMING Ask for critique explicitly. |
Creating homework for Periodic Table (Chemistry)