AIMentor — AI Expertise Learning Platform

Untimed. Answers are graded when you submit, with explanations and spaced-repetition scheduling.

Question 1easy
What is 'context engineering'?
Writing clever single promptsDeciding what the model can see, and in what order, within the finite context windowFine-tuning the modelChoosing the cheapest model
Question 2medium
What is the 'lost in the middle' effect?
The model forgets the system promptModels attend most reliably to the start and end of context; middle material can be under-usedTokens in the middle cost moreThe API drops the middle of long prompts
Question 3medium
Why is 'stuff everything into the prompt' often a bad strategy?
It always exceeds the context windowPadding relevant content with marginal context dilutes signal and raises cost/latencyThe API rejects large promptsIt disables caching permanently
Question 4hard
Prompt caching is a prefix match. What follows from that?
You can change any part of the prompt freely without losing the cacheAny byte change in the cached prefix invalidates everything after it — keep stable content firstCaching only works for the outputCaching requires temperature 0
Question 5medium
How do you verify a prompt cache is actually being hit?
The response is shorterusage.cache_read_input_tokens is greater than zero on repeat requestsThe model says sostop_reason is 'cached'
Question 6hard
Which silently break prompt caching by changing the prefix? (Select all that apply)
datetime.now() interpolated into the system promptA UUID placed early in the contentA tool set that varies per requestReusing the exact same system prompt every call
Select all that apply.
Question 7medium
For a long conversation approaching the context limit, what's a good strategy?
Keep appending every turn foreverSummarize/compact older turns and retrieve relevant history, keeping recent turns verbatimSwitch to a smaller modelLower the temperature
Question 8medium
What is compaction in the context of long conversations?
Compressing the model's weightsReplacing older context with a concise summary to free tokens while preserving continuityDeleting the whole historyReducing max_tokens
Question 9medium
How does 'memory across sessions' actually work?
The model retains everything internally between sessionsYou store facts externally and load the relevant subset back into context in a new sessionThe API keeps a hidden session cache foreverMemory requires fine-tuning
Question 10hard
In a multi-agent system, why give each sub-agent its own scoped context?
To make them slowerTo keep each window focused (better attention, lower cost) and enable parallel workBecause the API forbids sharing contextTo disable caching
Question 11easy
What is the typical render order that affects caching and attention?
messages → system → toolstools → system → messagesoutput → inputrandom each request
Question 12medium
Instead of stuffing a large knowledge base into every prompt, what's the scalable approach?
Increase max_tokensStore it externally and retrieve only the relevant pieces per query (RAG)Fine-tune on it every dayRaise the temperature
Question 13medium
One word: the caching technique that reuses an identical leading portion of the prompt across requests (a ____ match).
Question 14hard
One phrase/word: the principle that the RIGHT tokens matter more than MORE tokens (relevance over ____).

0 / 14 answered

Context Engineering — Practice quiz