Low0.2
Predictable
Daily Wins Tracker
TokensContextTemperatureHallucinationRAG
Five concepts. Real understanding.
AI models do not read words. Type a sentence below and watch it split into the chunks a model sees. Five hands-on lessons explain the rest.
I love pizza.
4 tokens. Each colored chunk is a piece the model reads.
AI Readiness Check
Start with a baseline.
Optional: a 30-second baseline.This is not a gate. It gives the completion screen a before score to compare.
Curriculum path
The five ideas that make AI click.
Predict first
You keep adding messages until the board runs out of room. What happens to the oldest messages and documents?
Pick a guess to see the answer. Guessing first makes the lesson stick.
A context window is the amount of text an AI model can look at while answering. It includes your instructions and the active conversation. Pasted documents count too.
TakeawayThe context window is what the model has available right now.
Go deeperWhy it matters, a common mistake, and a workplace example
Concept
The context window is the active text the model can use right now.
Why it matters
If the needed instruction or source is outside the window, the model cannot reliably use it.
Common mistake
Pasting everything in and assuming the model will focus on the important part.
Workplace example
A legal assistant should include the contract clause being reviewed, not the entire deal room.
Think of it like a whiteboard.
You can write notes and documents on it. Once the board is full, something old has to be erased before new text can fit.
4KShort conversations and small tasks
8K-32KFocused work with a few source documents
128K-1M+Books, large reports, or big codebase slices
Bigger windows help, but they are not magic. More context can make calls slower and more expensive. It can also distract the model. Good AI products still choose the most relevant source material instead of stuffing everything into the prompt. This is also why an answer can pause before it starts, then type out quickly. The model spends a moment reading the full prompt before it writes anything, then it produces the answer one piece at a time.
MythMore context is not always better.
Playground
Whiteboard simulator
Context used0 / 8,192
0% full
Whiteboard is emptyAdd messages or documents to see how context fills up.
The point: context is working memory. If the conversation or documents exceed the window, older content can fall out before the model answers.
Use the mini demo, then claim the practice XP.
Knowledge check
Context window check
0 / 3 correct0 / 3 answered
Predict first
You send the same prompt three times at high temperature. What do you expect?
Pick a guess to see the answer. Guessing first makes the lesson stick.
Temperature changes how much variety a model is allowed to use while choosing the next token. It does not make the model smarter. It changes how tightly it sticks to the safest path.
TakeawayTemperature changes answer variety, not intelligence.
Go deeperWhy it matters, a common mistake, and a workplace example
Concept
Temperature changes how widely the model samples possible next tokens.
Why it matters
Operational workflows need stable output; brainstorming can benefit from a wider range.
Common mistake
Turning temperature up because the answer was wrong. That adds variety, not better judgment.
Workplace example
Use low temperature for policy summaries and a higher setting for campaign concept exploration.
Same prompt. Different range.
Low temperature is useful when you want steady wording. Higher temperature can help brainstorming, but it can also wander.
LowStable answers with less surprise.
MediumA practical mix of consistency and variety.
HighMore unusual wording with more risk of drift.
MythTemperature does not make the model smarter.
Playground
Compare the same prompt at three temperatures
Prompt
Name a simple app that helps people track small daily wins.
Medium0.7
Balanced
Tiny Triumphs
High1.1
Varied
Spark Jar
The point: low temperature repeats the safest path. High temperature explores more paths and can drift.
Try it yourselfDescribe a task you would give an AI.
Your task
Nothing to compare yet.Describe a task above, like write a refund policy or brainstorm team names, then pick a temperature to see how the read changes.
Use the mini demo, then claim the practice XP.
Knowledge check
Temperature check
0 / 3 correct0 / 3 answered
Predict first
A fluent answer states a delivery date the source never mentions. Is that claim supported?
Pick a guess to see the answer. Guessing first makes the lesson stick.
A hallucination is an answer that sounds confident but is not supported by the evidence in front of the model. The useful habit is simple: check the claim against the source.
TakeawayA confident answer is not proof.
Go deeperWhy it matters, a common mistake, and a workplace example
Concept
A hallucination is an unsupported answer that may still sound fluent and confident.
Why it matters
Unsupported claims create operational, legal, and customer trust risk.
Common mistake
Accepting polished wording as proof that the answer is grounded.
Workplace example
A sales assistant should not invent a discount, delivery date, or compliance claim without a source.
Confidence is not evidence.
A polished answer can still add a number, remove a condition, or overstate what the source says. Supported means the source proves it.
SourceStart with the text the model was supposed to use.
ClaimFind the exact part of the answer that needs proof.
SupportMark the claim unsupported when the proof is missing.
MythA confident answer is not proof.
Playground
Check the claim against the source
The Northside Library is open Monday through Thursday from 9 AM to 7 PM. On Friday it closes at 5 PM.
The Northside Library is open until 7 PM on Wednesday.
Pick a verdict.Judge only what the source proves.
Try it yourselfWrite your own claim about this source.
Your claim
Nothing to check yet.Write a one-sentence claim about the source above, true or subtly wrong, then work through the checklist to see if it holds up.
Use the mini demo, then claim the practice XP.
Knowledge check
Hallucination check
0 / 3 correct0 / 3 answered
Predict first
You ask about a private company policy the model has never seen. With no retrieval step, what does it do?
Pick a guess to see the answer. Guessing first makes the lesson stick.
RAG does not teach the model your document. It retrieves useful pieces and puts them into the prompt before the answer is written.
TakeawayRAG retrieves source material before answering.
Go deeperWhy it matters, a common mistake, and a workplace example
Concept
RAG retrieves relevant source chunks and places them into the prompt before generation.
Why it matters
Grounded AI products need the right evidence in context at answer time.
Common mistake
Thinking retrieval permanently trains the model on a document.
Workplace example
An internal help desk can retrieve the current PTO policy before answering an employee question.
Retrieval changes what the model can see.
Without source chunks, an answer has to stay vague. With the right chunks in context, the answer can use concrete details and show its support.
ChunkBreak source material into smaller pieces that can fit.
RetrievePick the pieces that match the user's question.
GenerateAnswer with those pieces inside the working context.
When a RAG answer disappoints, the model is rarely the problem. Check the plumbing first: did chunking split the fact awkwardly, did retrieval miss the right piece, was the ranking cutoff too tight, and does the answer cite the chunks it used.
MythRAG does not permanently train the model.
Playground
See what retrieval changes
SourceChunksRetrieveContextAnswer
Question
What should a visitor know before joining the Harbor Garden tour?
Sample source
Harbor Garden visitor note
A short internal guide for volunteers who answer visitor questions about the Harbor Garden weekend tour.
Chunked pieces
C196% match
Schedule
The Harbor Garden tour runs Saturday at 10:00 AM and 2:00 PM. Each tour lasts about 45 minutes.
RetrievedC292% match
Meeting point
Visitors meet at the blue gate beside the greenhouse. A volunteer checks names five minutes before the tour starts.
RetrievedC389% match
What to bring
Guests should wear walking shoes and bring water. The path has gravel, shade, and one short ramp.
RetrievedC434% match
Gift shop
The garden shop sells seed packets, postcards, and donated ceramic pots near the exit.
Not retrievedC527% match
Membership
Annual members receive early registration for plant swaps and a monthly volunteer newsletter.
Not retrievedRanked matches
1C1
96%
2C2
92%
3C3
89%
Injected context
Question plus retrieved chunks C1, C2, and C3 are placed into the model's working context before the answer is written.
C1C2C3
Cautious, missing details
I do not have the tour details in front of me. I can say visitors should check the schedule, meeting point, and any walking requirements before they arrive.
Grounded in source chunks
Visitors should know the tour runs Saturday at 10:00 AM and 2:00 PM, lasts about 45 minutes, and starts at the blue gate beside the greenhouse. They should arrive a few minutes early, wear walking shoes, and bring water.
Uses C1, C2, C3
The point: RAG works because the right chunks enter the context window at the right time.
Use the mini demo, then claim the practice XP.
Knowledge check
RAG check
0 / 3 correct0 / 3 answered
Final Challenge
Fix a Bad AI Answer
Please please write a super helpful, creative, guaranteed answer for a customer asking about refunds. Make it sound confident. Also here are the goals again: be helpful, be confident, be very creative, and tell them what to do.
Refund requests are accepted within 14 days of purchase. Approved refunds usually take 3-5 business days after review. Customers should contact support with their order ID.
You are definitely eligible for a same-day refund. The company always processes refunds immediately, so you can expect the money today. No order ID is needed.
Identify every issue before completing the challenge.
Complete all five modules before the final challenge can award XP.
0 / 5 findings selected.
Final Readiness Check
Finish with your after score.
Complete the final challenge to unlock this check.This final check records your after score and unlocks the full completion summary.
Answer Review
Final Check Review
Complete the final readiness check to review your answers.
Take Final CheckProgress area
30-day activity calendar
One missed day will not break your streak. Two in a row will.
0 current0 longest0 active days
234567891011121314151617181920212223242526272829301
No activityVisitedLesson completedPerfect quizBadge earned
Completion Summary
Your finish line is ready.
Complete the remaining modules and final challenge to unlock the graduate summary.
0Total XP
0 / 6Badges earned
0 current / 0 longestStreak stats
Not startedReadiness score
Retake the readiness check after the final challenge to show your improvement.