Practicum glossary
The working minimum: the terms the worksheets actually use, in plain words. This is not the glossary of the books — each book has its own, fuller one. Here is just enough to fill a sheet in and know what it is asking.
If you work through the Practicum with an AI assistant, hand it this page along with the material. Then it uses these meanings rather than whatever it picked up on the open web.
Value
Output. The text, report, layout, code or deck — the thing you can hand over. Output is what AI made cheap: almost anyone can now produce it, fast. Where it works: the three value levers.
Outcome. Not "made the deck" but "the client understood and bought". Not "wrote the policy" but "people work by it". An outcome cannot be handed in — someone is accountable for it. The core shift of the series: price moves from output to outcome.
Value lever. The way your work turns into money. There are three: ① growth and new revenue (what others cannot do), ② cost saving (the same thing cheaper or faster), ③ attention and trust (the reason people choose you). A task that pulls none of them is busywork. Where it works: the three value levers.
Busywork. Work that fills the day and pulls no lever. Moving numbers between spreadsheets; a report nobody reads. More dangerous than idleness, because it feels like being busy.
Transferability. Whether the value you create can exist without you personally being there. Low transferability means "it works only while I am at the desk" — advice that cannot be handed on. High means a method, a template, a product, a team you trained. The lower it is, the harder your income ceiling: it is capped by your hours. Where it works: the weak multiplier canvas.
Exposure. What share of your working time a machine can already reproduce — the time in baskets 🟢 and 🟡. Not a sentence; a map of where you have slack and where you do not. Where it works: the week inventory.
Tasks and levels
The three baskets. A way to sort a week. 🟢 "AI on its own" — routine work with a clear input and a checkable output. 🟡 "human + AI" — the machine gives volume or a draft, the human holds context and the check. 🔴 "human only" — judgment, accountability, trust, difficult conversations. Baskets drift; redo this about every six months.
The five levels of usefulness. A ladder from executor to system-builder: 0 — follows instructions, 1 — uses AI, 2 — validator (checks quality and risk), 3 — orchestrator (assembles a process from people, agents and data), 4 — owner of the outcome (accountable for a metric, money, a client), 5 — builds the system (it works without you). A level describes a task, not a person: you can be a 4 in one task and a 1 in another. Where it works: level of usefulness.
The five deficits. What grows more valuable as routine work gets cheap: judgment, taste, trust, attention, distribution. A deficit can be uncounted — you have it, but the market or your manager does not see or pay for it. That is a different problem, and it needs a different move than "build the skill further". Where it works: the five deficits audit.
Working with AI
Prompt. A request to an AI, in words. A good one carries the role, the input, the steps, the answer format and the point where you check. A bad one asks it to "make it nice". Where it works: the prompt method.
Harness. Everything you have built around the AI so it returns a steady result: how you frame a task, what you feed it, where the context lives, how you check, where you keep what worked. The model is much the same for everyone — the harness is the difference. Where it works: your personal system, examples by level.
Skill (SKILL.md). An instruction file for an AI agent: when to apply it, what goes in, the
steps, what comes out and how it is checked. The agent picks it up on its own when a task falls into
its area. Simple rule: explaining the same logic for the fifth time means it is already a skill.
Where it works: turn a habit into a skill,
ready-made skills.
Agent. An AI-based program that does not just answer but carries out steps: searches, writes, sends, reaches into other systems. It acts — which is why it needs boundaries and an owner.
Cognitive debt. A skill you quietly lose because the AI performs it. It shows up late: first "I just got faster", then "it is hard to put this into words myself". Three common modes: cyborg (AI amplifies, you hold the meaning), centaur (you split the work deliberately), self-automator (you hand it over and stop reading). Where it works: the cognitive debt checklist.
Business and agents
AI-native. A process built around what AI can do from the start, rather than an old process with a model bolted on. The test: remove the AI — does the process fall back to its old shape, or does it break entirely? Where it works: the process maturity test.
Agent contract. The agreement on what an agent does and does not do: role, inputs, tools, memory, boundaries, escalation, review, metrics and a human owner. Without boundaries, escalation and an owner it is a wish, not a contract. Where it works: the agent contract.
Gate. The point where a change either passes or stops — named in advance, before launch, not after an incident. A gate has a metric, a threshold number and an action at that threshold. Where it works: the change gates sheet.
Eval set. Real tasks with known-correct answers, used to test an agent before it fails in the open. Without one, "it works" is a feeling, not a fact. Where it works: the eval set builder.
Shadow mode. The agent runs alongside a person, but its decisions are not applied — they are compared. A way to learn the cost of an error without paying it.
Agent customer. A buyer who arrives through an AI rather than in person: the assistant searches, compares and sometimes purchases. It needs something different from you — to be findable, assessable and buyable without a call. Where it works: visibility in AI search.
Platform rent. The share you hand a platform for access to customers, data or tools. It grows unnoticed and becomes visible once leaving is expensive. Where it works: the platform rent map.
Trust
Cost of being wrong. What you lose if you believe something false. This sets the depth of checking — not how convincing the message looks. The other side controls how convincing it is. You control the cost. Where it works: the trust compass.
Second channel. A way to confirm that the sender does not control. Not "call the number in the message" but "call the number I already had".
Verifiability. The ability to confirm a claim independently. Different from transparency: transparency is "we explained how it works", verifiability is "you can check for yourself".
Chain of custody. The trail showing where a file or record came from and what happened to it. It is assembled in advance; after a dispute starts it is too late. Where it works: chain of custody.
Synthetic. Made by AI and presented as genuine: text, voice, face, review, candidate, customer. The word is neutral — synthetic is not automatically harmful, but it calls for a named check.
A term that is not here is explained inside its own worksheet — each sheet stands on its own. Fuller glossaries with examples live in the books themselves, one per volume.