Hone · 60-minute workshop

Class glossary

A glossary for briefing, checking, and reusing evidence-backed AI work.

No dumb questions. Use this page during the workshop and keep it afterward.

Terms
26
Levels
6
Resources
5
00
easy

AI basics

10 terms

The small set of AI terms every workshop builds on.

Artificial intelligence (AI)#

Software that performs tasks associated with human intelligence, such as understanding language, finding patterns, making predictions, or choosing actions.

Why it mattersIt is the broad category. Models, language models, and agents are specific kinds of AI systems.

You’ll hearWe use AI to summarize the feedback.

Model#

A trained mathematical system that turns input into a prediction or output.

Why it mattersThe model is the engine that produces an answer. The surrounding product decides what context and tools it receives.

You’ll hearWhich model is this using?

Large language model (LLM)#

A model trained on large amounts of text to predict and produce language one token at a time.

Why it mattersLLMs power chat assistants and many agents, but they still generate predictions rather than retrieving guaranteed facts.

You’ll hearSend that prompt to the LLM.

Token#

A small unit of text that a language model reads or writes. A token may be a word, part of a word, or punctuation.

Why it mattersContext limits, speed, and API prices are measured in tokens.

You’ll hearThat document uses too many tokens.

Prompt#

The instruction or input given to a model.

Why it mattersA clear prompt states the desired outcome, useful context, constraints, and how to recognize a good result.

You’ll hearTry the task again with a clearer prompt.

Context window#

The finite amount of instructions, conversation, source material, and tool results a model can consider at one time.

Why it mattersImportant information can be omitted or pushed out when the context becomes crowded.

You’ll hearThat conversation filled the context window.

Try it: Fill a context window
Try it · interactive illustration · no AI call

A context window has edges.

Try adding source material to this 12-unit window. Keep two units for instructions and two for a response.

Instructions: 2Response space: 2Sources: 6
10 of 12 units used. 2 left.

Illustrative units, not real token counts. Models and tools have different limits and may reject, truncate, or summarize long inputs. A context window is not permanent memory.

Hallucination#

A fluent answer that contains invented or unsupported information.

Why it mattersConfident wording is not evidence. Important claims still need a source or another check.

You’ll hearVerify that citation in case the model hallucinated it.

Try it: Trace a claim to its source
Try it · interactive illustration · no AI call

Does the source actually support it?

Choose a claim, read the passage, then make the call. A citation alone is not proof.

Claim to checkThe workshop fee is $800.
Fictional training proposal · Fees

Workshop fee: $800. Travel is extra; a travel quote has not been supplied.

Read the source before deciding.

Grounding#

Giving a model trusted source material and asking it to base its answer on that evidence.

Why it mattersGrounding makes answers easier to verify and reduces unsupported claims.

You’ll hearGround the answer in the workshop materials.

Try it: Trace a claim to its source
Try it · interactive illustration · no AI call

Does the source actually support it?

Choose a claim, read the passage, then make the call. A citation alone is not proof.

Claim to checkThe workshop fee is $800.
Fictional training proposal · Fees

Workshop fee: $800. Travel is extra; a travel quote has not been supplied.

Read the source before deciding.

Agent#

A model operating in a loop with instructions and tools so it can take multiple steps toward a goal.

Why it mattersAn agent can inspect results and continue working instead of stopping after one response.

You’ll hearGive the agent the goal and let it run the checks.

Try it: Walk through an agent loop
Try it · interactive illustration · no AI call

Walk through one bounded agent loop.

Advance the illustration yourself. Nothing here calls a model, uses an external tool, or spends money.

  1. 1Decide
  2. 2Act
  3. 3Check
  4. 4Return and stop

Work out which supplied evidence is needed for the assignment.

Tool use#

An agent calling an external capability such as reading a file, searching the web, running code, or sending a request.

Why it mattersTools let a model affect systems and gather current evidence, which also means permissions and limits matter.

You’ll hearThe agent needs a tool for that action.

Try it: Follow a tool call
Try it · interactive illustration · no AI call

A tool call is a round trip.

Advance the illustration yourself. Nothing here calls a model, uses an external tool, or spends money.

  1. 1Request
  2. 2Tool result
  3. 3Use the result

The model asks an authorized calculator to evaluate 800 × 1.05.

01
easy

Define the work

4 terms

Give the assistant a clear assignment and rules for making the comparison.

Bounded task#

A specific piece of work with a defined result and limits on what the assistant may do. In this workshop, AI compares options and recommends; the person decides.

Why it mattersYou can delegate useful work without granting permission for purchases, approvals, or other outside actions.

You’ll hearHand off the work, not the decision.

Handoff brief#

Instructions naming the decision, requirements, available evidence, work to perform, checks, and stopping point.

Why it mattersYou can remove important guesswork before asking the assistant to complete the task.

You’ll hearNothing important left to guess.

Try it: Explore a complete handoff
Try it · interactive illustration · no AI call

Change the task, not the structure.

Choose a part to see what it contributes. These examples explain the structure; use the guide’s full prompt for the workshop.

Decision

Compare two training proposals for our team.

AI returns the workYou check and decide

Hard limit#

A requirement an option must satisfy to remain eligible, such as a maximum total cost or a policy restriction. Attractive features cannot override it.

Why it mattersYou can rule out unsuitable options before weighing their advantages.

You’ll hearThis is a hard limit, not a preference.

Try it: Test a hard budget limit
Try it · interactive illustration · no AI call

Watch tax change the comparison.

A sticker price is not an all-in cost. Add the assumed tax before deciding what fits.

Limit: $1,000.00Before tax: not the final comparison
Proposal A$800.00 + $0.00 tax = $800.00
Add tax before judging eligibility
Proposal B$1,100.00 + $0.00 tax = $1,100.00
Add tax before judging eligibility
The full budget comparison is not ready until tax is included.

Fictional complete-price proposals with an assumed 5% tax, only for this illustration. Not changing anything remains an option; transition costs would need separate evidence.

Priority order#

The ranking of preferences used to compare options that meet the hard limits. A higher-ranked preference matters more, but does not authorize invented scores or evidence.

Why it mattersYou can explain what should guide the recommendation when eligible options involve trade-offs.

You’ll hearWhat matters most?

02
easy

Establish the evidence

4 terms

Separate what the sources support from assumptions and missing information.

Source packet#

The supplied material the assistant is allowed to use as evidence. The workshop uses a fixed, dated text snapshot with source links and qualifications, not a complete or current answer to every question.

Why it mattersYou can give the assistant readable evidence and check its answer against the same material.

You’ll hearUse only the supplied sources.

Permitted assumption#

An input explicitly authorized for the task, such as an assumed rate used in calculations. It must remain distinguishable from a fact established by a source.

Why it mattersYou can allow necessary calculations without letting the assistant quietly invent favorable conditions.

You’ll hearShow which inputs are assumptions.

Citation#

A pointer to the passage supporting a factual claim. In the workshop, a heading and short supporting phrase help you locate and inspect that evidence.

Why it mattersYou can test whether a claim is supported instead of trusting a reference that merely mentions the topic.

You’ll hearDoes the passage actually support it?

Try it: Trace a claim to its source
Try it · interactive illustration · no AI call

Does the source actually support it?

Choose a claim, read the passage, then make the call. A citation alone is not proof.

Claim to checkThe workshop fee is $800.
Fictional training proposal · Fees

Workshop fee: $800. Travel is extra; a travel quote has not been supplied.

Read the source before deciding.

Material unknown#

Missing information that could change the recommendation. The assistant should explain its effect rather than fill the gap from memory or treat it as a verified benefit.

Why it mattersYou can identify what needs more evidence and decide whether a conditional recommendation is usable.

You’ll hearWhat could change the choice?

Try it: Trace a claim to its source
Try it · interactive illustration · no AI call

Does the source actually support it?

Choose a claim, read the passage, then make the call. A citation alone is not proof.

Claim to checkThe workshop fee is $800.
Fictional training proposal · Fees

Workshop fee: $800. Travel is extra; a travel quote has not been supplied.

Read the source before deciding.
03
medium

Define done and where to stop

3 terms

Specify the finished work while keeping authority with the person.

Strongest objection#

The most important reason against the recommended option, including a need it leaves unmet or evidence that could favor another choice.

Why it mattersYou can inspect the recommendation's main weakness instead of receiving only reasons to agree.

You’ll hearWhat is the strongest reason against this recommendation?

Stopping point#

An explicit boundary telling the assistant when to stop and which actions are outside its authority. It should also stop and explain a blocker if a necessary input or contradiction prevents completion.

Why it mattersYou can request a recommendation without authorizing sending, purchasing, publishing, or approving.

You’ll hearReturn the finished work, then stop.

04
medium

Inspect before accepting

3 terms

Check the actual result, then accept it or request a focused repair.

Acceptance checks#

The tests you use to decide whether the work is usable: required options are covered, claims have evidence, costs are correct, the conclusion respects requirements, and uncertainty remains visible.

Why it mattersYou can accept a brief based on inspected evidence rather than the assistant's confidence.

You’ll hearCheck the work, not the confidence.

Try it: Check a confident answer
Try it · interactive illustration · no AI call

Check the arithmetic, not the confidence.

This deliberately incomplete practice example left out tax. It is not a claim that your assistant made this mistake.

Illustrative answer“I checked it. The total is $800.00.”
Your independent check$800.00 + ($800.00 × 0.05) = ?
Calculate it yourself, then choose the total.

Passing one calculation is not approval of the whole brief or permission to buy. Still check evidence, coverage, requirements, and judgment.

Independent verification#

Checking the output directly against sources and separately recalculating important numbers, rather than asking the same assistant whether it was right.

Why it mattersYou can catch unsupported claims and arithmetic errors even when the assistant says it verified everything.

You’ll hear“I verified it” is a claim, not proof.

Try it: Check a confident answer
Try it · interactive illustration · no AI call

Check the arithmetic, not the confidence.

This deliberately incomplete practice example left out tax. It is not a claim that your assistant made this mistake.

Illustrative answer“I checked it. The total is $800.00.”
Your independent check$800.00 + ($800.00 × 0.05) = ?
Calculate it yourself, then choose the total.

Passing one calculation is not approval of the whole brief or permission to buy. Still check evidence, coverage, requirements, and judgment.

Bounded correction#

A focused follow-up naming an observed defect, the required fix, and the check it must pass while preserving unrelated requirements. Any affected calculations and conclusions must be updated too.

Why it mattersYou can repair a real problem without restarting the assignment or entering an endless retry loop.

You’ll hearName the defect and the acceptance check.

05
advanced

Test and reuse the method

2 terms

Change an input deliberately, preserve the other rules, and save what made the work checkable.

Controlled rerun#

Repeating the assignment after changing one requirement while keeping the other inputs fixed. The dependent calculations and reasoning must update, even if the recommendation stays the same.

Why it mattersYou can test whether the instructions still work when circumstances change without allowing silent changes to other rules.

You’ll hearPredict first. Then rerun.

Try it: Change one input
Try it · interactive illustration · no AI call

Move one limit. Keep the other rules.

Predict which options will fit, then move the budget. Prices, sources, and the assumed tax rate stay fixed.

Limit: $1,000.00Including 5% tax
Proposal A$800.00 + $40.00 tax = $840.00
Fits this estimate
Proposal B$1,100.00 + $55.00 tax = $1,155.00
Over budget
1 of 2 proposals fits at these estimates.

Fictional complete-price proposals with an assumed 5% tax, only for this illustration. Not changing anything remains an option; transition costs would need separate evidence.

Sources fixedPriorities fixedNo invented discount

Reusable handoff#

A saved method that keeps the instructions, dated sources, checked result, unresolved questions, and acceptance checks together. Changed versions are labeled so their conclusions are not confused.

Why it mattersYou can repeat or adapt the work while preserving the evidence, limits, and checks that made it usable.

You’ll hearSave the method, not just the answer.

Try it: Explore a complete handoff
Try it · interactive illustration · no AI call

Change the task, not the structure.

Choose a part to see what it contributes. These examples explain the structure; use the guide’s full prompt for the workshop.

Decision

Compare two training proposals for our team.

AI returns the workYou check and decide
Keep going

Workshop resources

Hand Off a Real Task to AI, Start to FinishNick Nisi · Updated 2026-09-22