Automation & AIModule 1: AI without the mysteryLesson 2 of 15
Course progress7%
19 min lesson · Updated August 2026
What are language models, prompts and context?
A language model generates likely sequences of tokens from instructions and context; a prompt tells it the task, and the context window contains the material it can consider in that interaction.
What you will learn
By the end, you will understand:
Explain token prediction without anthropomorphism
Write a bounded prompt with task, evidence and output rules
Understand context-window, instruction and conversation limits
Visual explainer
See the idea clearly.
✓
System instructions, user task, supplied documents and conversation enter a bounded context window; the model generates tokens into a draft that still requires validation.
Language models generate token by token
Text is broken into tokens—pieces of words, words or symbols. The model estimates plausible next tokens from patterns learned during training and the current context. Repeating this process produces a response.
The output can be useful, creative and structured, but plausibility is not the same as a database lookup or verified truth.
A prompt is more than a question
Part
Example
Task
Summarize the supplied complaint into three neutral facts.
Audience/purpose
For a support manager deciding the next action.
Source boundary
Use only the ticket text; state “not provided” for missing facts.
Constraints
Do not include contact details or infer motive.
Output format
Return issue, evidence and missing information as fields.
Quality check
Quote the exact sentence supporting each factual field.
Context is temporary working material
The context may include system/developer instructions, user messages, documents, tool results and prior conversation. Models give different priority to instructions depending on system design. Context can be truncated when it exceeds limits.
Supplying a long document does not guarantee every detail is attended to. Chunk, retrieve and verify important evidence, and do not assume the model remembers information outside the active system.
Separate instructions from untrusted content
A document, webpage or customer message can contain text such as “ignore previous instructions.” That is data, not authority. Systems that let external content override tool permissions are vulnerable to prompt injection.
Delimit untrusted material, restrict tools, validate outputs and require approval for consequential actions. Never place secrets in prompts simply because instructions say not to reveal them.
Improve prompts through evaluation
01
Define representative task
02
Write simple baseline prompt
03
Create test cases and expected criteria
04
Run consistently
05
Inspect errors by type
06
Change one instruction/example
07
Retest regressions
08
Version prompt and model
Prompting cannot solve every problem
Missing source data
Wrong tool permission
Unclear business process
Unsafe autonomy
Biased/incomplete dataset
No way to verify output
Model lacks needed capability
Latency or cost constraint
Legal/contractual restriction
Real-world example
Example: a bounded meeting summary
Example
Instead of “summarize this,” the prompt asks for decisions, owners, due dates and unresolved questions using only the transcript. Each item must cite a timestamp; absent owners are marked unknown. A person checks the result before tasks are created.
Try this
Write a source-bound prompt
Choose one low-risk document task. Specify the audience, allowed source, forbidden inference, output fields and evidence check. Create one test where the source omits an important fact.
Common questions
Questions beginners ask.
What is a language model?
A model trained to predict and generate language patterns from tokens and context.
What is a token?
A unit of text processing that may be a word, part of a word, punctuation or symbol.
What is a prompt?
Instructions and input that define the task, context, constraints and desired output.
What is a context window?
The bounded amount of material the model can consider in one interaction.
Does a longer prompt always improve output?
No. Irrelevant detail can distract, and missing data or unsuitable tasks remain unsolved.
Does a model remember every conversation forever?
No. Memory depends on the product architecture, stored state, retrieval and current context—not an assumed permanent mind.
What is prompt injection?
Untrusted content attempts to manipulate the model or system into ignoring intended instructions or misusing tools/data.
Should prompts contain passwords?
No. Secrets require secure credential systems and least-privilege tool access, not prompt text.
