Every lesson and reference page uses these terms and no synonyms. The “avoid” lines are the point — several of them mark wordings this curriculum used in an earlier revision and had to correct.
Canonical language for this workspace. Every lesson, reference document and quiz item uses these terms and no synonyms. The Avoid lines are the point — they name the vaguer or wronger word each term replaces, and several of them mark mistakes this curriculum made in an earlier revision and had to correct.
Terms are added as they become load-bearing.
Generation: Producing an answer sequentially, one chunk of text at a time, each shaped by everything in front of it. The root mechanism the whole curriculum hangs off. Avoid: Thinking, computing, working out the answer
Chunk (token): The unit text is handled in — not letters, not quite words. Named once in Lesson 1 because learners meet the word "token" elsewhere, then dropped. Avoid: Word, character, piece
Unaided generation: Producing an answer with no supplied document and no retrieval tool in play. The condition under which fabrication is most likely. Avoid: Reconstruction, recall, remembering
Deliberate correction. An earlier revision taught "it reconstructs, it does not retrieve," which collapses the moment Projects, chat search, connectors and web search all retrieve. Claude retrieves constantly. The distinction is whether this answer was retrieved or generated.
Cutoff: The point after which the model learned nothing. It generally cannot tell that it doesn't know, and asking it its own cutoff is not a reliable check. Avoid: Training date, last update, knowledge limit
Sampling: Why the same prompt twice can give two different answers. The reason rerolling an unchanged prompt is not a repair. Avoid: Randomness, temperature, variation
Available information: Everything Claude can reason over for the response it is about to give — this conversation, what you attached, and whatever memory, project knowledge, chat search, connectors or web search brought in. Avoid: The context window, what it remembers, what it knows
Deliberate correction. "Context window" names a container and invites the false claim that anything outside it does not exist. Several mechanisms cross that boundary.
Scope: How widely a piece of persistent context applies — everywhere (Instructions for Claude), one body of work (a Project), or accumulating on its own (memory, which itself has a per-project scope). Avoid: Bucket, layer, container, box
Grounding: Attaching a claim to something inspectable. Makes a claim traceable; does not make it true. Avoid: Fact-checking, sourcing, making it accurate
Traceability: Whether you can find what a claim rests on. One of the two independent questions in Lesson 5 — the other is whether that source is any good. Avoid: Reliability, accuracy, trustworthiness
Specification: What a prompt actually is. The six components are task, context, input, output shape, constraints and non-goals, and examples. Avoid: Prompt engineering, prompt craft, magic words
Failure class: Which of the six named things went wrong: missing information, under-specified, fabrication, wrong shape, lost the thread, insufficiently decomposed. Avoid: Bad output, it didn't work, hallucination (as a catch-all)
Insufficiently decomposed: A task handed over whole that needed breaking into steps. Named for the repair, not for a supposed capability ceiling. Avoid: Task too big, beyond its abilities
The reroll trap: Regenerating an unchanged prompt and expecting a different class of answer. Draws again from the same distribution. Avoid: Trying again, retrying
Diagnosis-as-hypothesis: What you get when you ask Claude why it produced something. Useful for generating leads; not a log of what actually happened. Avoid: Explanation, introspection, the reason it did that
Tier A / B / C: The robustness test every exercise must pass. A — analysis of a supplied artifact, cannot break. B — the learner constructs something and inspects it, no required outcome. C — depends on a specific model behaviour, and does not ship. Avoid: Static exercise, live exercise, demo
Supplied artifact: A captured, dated, real output that a lesson is built around, so the teaching event has already happened and cannot fail to recur. Avoid: Example, sample, screenshot
Volatile specific: A fact that will change and that a learner does not need in order to act — a page threshold, an effort-tier name, a model lineup. Cut on sight. Avoid: Detail, spec, parameter