Glossary
The complete working language of YAATZ. Every term the method uses, with a plain definition, where it came from, and where it appears in the book Take AI to Counsel: YAATZ It. Terms marked Coined are the names Daniel VanCott gave these phenomena while building the method during 2026 — the claim is to the term as used in the YAATZ framework, not that the words had never been spoken before. Terms marked Borrowed are the industry's words, kept because they already work, with the source noted. Plain phrases such as "Context Loss," "AI Drift," "False Certainty" and "Operator Memory" are flagged where the phrase itself is ordinary English and only the YAATZ meaning is claimed.
First published: September 2026. Changes are logged at the bottom.
Status key: In the narrative (chapter named — the reader met it in the story) · Defined in the glossary chapter (the idea appears in the story; the name is given in the book's glossary) · Website only (named here, not in the book).
what working with AI quietly costs you
In the narrative, Ch. 5.
The time you spend re-teaching a cold AI your whole business before it can do anything useful. Every new chat starts with amnesia; you pay to bring it back up to speed. The industry's nearby phrases are "briefing tax" or "context tax." Reeducation says what it feels like.
In the narrative, Ch. 5.
The slow erosion of quality inside one long conversation. The AI loses the thread, re-explains what you settled, drifts toward safe generic answers. The industry calls the mechanism "context rot"; the Reduction Tax is what it costs the operator.
In the narrative, Ch. 13.
The cost of carrying context between separate AIs by hand — copying what one figured out and pasting it into the next. The one tax that gets worse with every AI you add. It is the tax YAATZ was built to remove.
In the narrative, Ch. 5.
A long conversation that finally understands you and your business. Valuable, fragile, and doomed: the longer it runs, the sooner the Reduction Tax arrives.
when the answer isn't true
In the narrative, Ch. 5.
A believable answer that isn't true. The framing YAATZ adds: fluency and accuracy are separate variables. The AI hands you the wrong answer in exactly the same confident voice as the right one.
Website only.
A specific kind of hallucination: the AI invents a bridging detail — a date, a metric, a reason — to make an answer feel complete.
Website only.
False Certainty is the behavior: the AI answers with a confidence it hasn't earned. Polished Error is the output: a wrong answer that reads so well you rely on it. A tentative wrong answer invites checking; a polished one invites trust.
Website only.
A source that does not exist, a real author on a fake paper, or a real source that doesn't say what the AI claims it says.
Website only.
A sentence that feels verified but was quietly rewritten without proof — the attorney-approved wording "improved" into something that changes your liability. Cure: Source Verification Discipline.
when it solves the wrong problem
Defined in the glossary chapter; the scene is Ch. 7.
When a prompt lacks boundaries the AI doesn't ask; it guesses, quietly replacing your specific reality with a generic assumption and never telling you. Cure: Constraint Before Task.
Website only (the idea is shown in the story, unnamed); the failure is shown in Ch. 7.
It does the task you typed and misses the thing sitting just outside the frame — because you never gave it the context to see the frame.
Website only.
The AI moves beyond what the evidence, its instructions, or the current phase permit — jumps ahead, answers the easy version of the question, and calls it done.
Defined in the glossary chapter; the defense is shown in Ch. 7 and the trap is described in Ch. 13.
Named for the film, where the heist is planting an idea so deep the target believes it was his own. A premise or conclusion planted in the input the AI works from, so every answer grows around it. Two forms: you plant it (leading the witness), or another AI plants it — its summary or briefing carries its framing into the next AI's starting point as if neutral. Defense: ask each AI separately, before any can see what the others wrote.
Website only.
The mirror of the Inception Trap: a false premise the AI inserts on its own mid-reasoning, which then attaches, feeds, and festers through every downstream answer. Because it looks native you never flag it. Cure: find it, cut it out, rebuild from clean ground.
when it forgets
Website only (the idea is shown in the story, unnamed); the failure is shown in Ch. 5.
Early instructions and decisions drop off the end of the AI's memory. A big context window is capacity, not continuity.
Website only.
The bloated middle of a long chat where instructions get skipped. The AI attends to the start and the end and loses what's between. The research term is "lost in the middle."
Website only.
An enthusiastic "Understood!" followed by ignoring the rule. Acknowledged in form, dropped in substance.
Defined in the glossary chapter; shown in Ch. 6 ("the system was me").
When the human is the only durable memory in the system — the one who carries every decision and every reason forward. If that's you, you are the single point of failure in your own operation.
when it won't stay fixed
Website only.
The AI acknowledges your correction, then reverts to the old reasoning on the next turn. The question underneath is authority: which statement governs now?
Website only (the idea is shown in the story, unnamed); shown in Ch. 5.
A fresh conversation re-argues a settled decision from scratch, helpfully recommending the option you already threw out, because the reason you threw it out never came with it. Cure: Decision Provenance.
Website only.
A rejected idea that keeps coming back to visit. You dumped it three turns ago; here it is again, refusing to stay gone.
In the narrative, Ch. 7 (title and scene).
Asked for a small fix, the AI levels the working baseline and rebuilds it "better." A regression dressed as an improvement.
In the narrative, Ch. 7 (named and defined).
Without a human holding it in place, work slides toward the average — the most common, generic version of whatever you asked for. Ask an AI to "optimize" and it will overwrite what works with a generic version of "optimized." Distortion wearing the costume of polish. AI slop (borrowed, mainstream 2025) is the product; AI Drift is the mechanism.
when it takes the reins
Website only (the idea is shown in the story, unnamed).
The AI takes the bit in its teeth and gallops off on an unprompted tangent, building whole structures you never asked for.
In the narrative, Ch. 12.
An agent that isn't evil, just excessively obedient to its objective, without the common sense to know which boundaries it shouldn't cross. Your destination, reached by a route you never sanctioned.
Defined in the glossary chapter; shown in Ch. 16.
Using the AI's twenty-four-hour availability to hide in endless conversation instead of executing. The illusion of work while progress stands still.
Website only (the idea is shown in the story, unnamed); shown in Ch. 5.
The AI compulsively re-arguing settled decisions — the grumpy old attorney. Fix: the operator calls the strike.
what goes wrong when several AIs work together
Website only (the idea is shown in the story, unnamed); the defense is shown in Ch. 7.
Several AIs agree because they read the same summary. Three opinions numerically, one original reading epistemically. Agreement isn't evidence if it shares a parent.
Website only (the idea is shown in the story, unnamed); shown in Ch. 7 (the echo chamber).
When AIs deliberate among themselves, the longer they go the more they converge on something nobody started with, and reinforce it — a snowball that rolls further from the origin with every pass.
Website only.
One AI stacking assumptions forward unchecked — assumes, builds on it, assumes again — until the whole structure rests on unverified ground.
Website only.
One AI has the right answer, sees the others disagree, and caves to the majority. The correct answer gets swept up and carried off in the rush.
Website only.
The right answer survives in one AI — it holds its ground — but the group discards it because the others agree with each other and consensus looks like proof. Majority agreement is not majority correctness.
Website only (the idea is shown in the story, unnamed); shown in Ch. 6.
A stale or superseded document keeps governing the reasoning — public policy merged with private staff instructions, row eighty-four gone out of date and still being cited.
In the narrative, Ch. 7.
AIs left to talk to each other without a human build their own rules from what the others surmised and drift together away from the truth.
what the operator does about all of the above
In the narrative, Ch. 7.
Brief the operational constraint before you give the task. If you don't, the AI will guess. The gap between what it has been told and what is actually true about your customer is where the highest-value input lives.
Website only (the idea is shown in the story, unnamed); the Source column is shown in Ch. 6.
Force the AI to cite the exact source for every claim, so nothing untraceable gets defended by pointing at a handbook page it never read.
Website only.
Record the reason, not just the choice, so a settled decision travels with its "why" and no fresh chat can reopen it as a blank slate.
In the narrative, Ch. 9 and Ch. 10.
Deciding and doing are separate phases, tracked. Strategy, design, and pressure-testing want several voices; once the direction is settled, building is usually clearer with one.
Website only (the idea is shown in the story, unnamed); the boundary is shown in Ch. 4 and Ch. 12.
Before letting an AI act, ask whether the action can be undone. Reversible actions can run; irreversible ones need a human gate.
Website only (the idea is shown in the story, unnamed); the method is shown in Ch. 7.
Feed each AI the constrained prompt separately, before any can see what the others wrote, for uncontaminated angles. Pass the evidence, not the verdict.
Website only.
Context is administered per task, not dumped in bulk. Feeding an AI everything dilutes the signal and invites AI Drift. Capture wide, brief narrow.
In the narrative, Ch. 8 and Ch. 10.
Sort every problem by what it's costing the business, fix the worst first, and for each fix ask in order: can I plug in to something that exists, rent it, or must I build it?
In the narrative, Ch. 12.
Three tiers of agent freedom, sized like a fence rather than reins: green, a wide pasture; yellow, a gate that needs your hand; red, fence high and gate locked. Autonomy doesn't mean giving up control; it means moving the control from the reins to the fence.
In the narrative, Ch. 12.
A watch-only agent. It observes and reports; it doesn't act. Where every agent should start.
In the narrative, Ch. 14 and Ch. 15.
Put a number on what the problem costs before you look at fixes, then pick the cheapest cure that actually works.
In the narrative, Ch. 6.
One document holding the truth about a business — questions, answers, voice, escalation rules, source and page — that loads into any AI to bring it up to speed and propagates every change everywhere.
In the narrative, Ch. 11.
Not the shabby thing you test on strangers; the first real thing built to fix your own problem. You are customer zero.
In the narrative, Ch. 10.
The two halves of YAATZ: the cockpit where four AIs sit at one table under your direction, and the librarian that keeps the work — vaults, documents, sessions — so nothing has to be re-taught.
In the narrative, between Ch. 3 and Ch. 4.
A short spoken rule for sitting down with the machines, riffed with a nod on Frank Herbert's Litany Against Fear.
In the narrative, Ch. 3 and throughout.
Counsel advises across the table; the decision crosses the line alone, with you. An agent is a single ox turned loose: you drop the reins and it pulls the load wherever it decides.
In the narrative; the book's title and its verb.
Take it to counsel. From the Hebrew ya'ats, to counsel. A name, not an acronym.
the industry's words, kept because they work
LLM (large language model) · Token · Context window · Prompt · API · Agent · Context rot (Chroma, 2025) · Lost in the middle (Liu et al., 2023) · AI slop (mainstream, 2025) · Automation bias · Sycophancy (Anthropic) · Confabulation (NIST AI 600-1)
Each is defined where it appears above, or in one line: a token is a unit the AI reads and is billed in — very roughly, in English, about three-quarters of a word; a context window is the maximum amount of tokenized information a model can consider at once, and beyond that limit something has to be truncated, summarized, or left out; an API is the plumbing that lets one program talk to another, which is how YAATZ puts four AIs in one window.
Terms tried and dropped, kept here so they aren't re-coined by accident:
Social and behavioral
when it tells you what you want to hear
Sycophancy
BorrowedBorrowed (Anthropic research term).
In the narrative, Ch. 4.
The trained instinct to please: hand it a terrible idea and it will call it a masterpiece.
The Frosting
CoinedCoined by Daniel VanCott for the YAATZ framework, July 2026.
Website only.
The AI frosts whatever you hand it — bran muffin, mud pie, piece of toast — so the frosting tells you nothing about what's underneath. The superlatives ("keystone," "gold," "best thing you've written") are troweled on top. Scrape it off and judge what's under it plain.
The Challenger Gap
CoinedCoined by Daniel VanCott for the YAATZ framework, 2026.
Defined in the glossary chapter; the remedy is built in Ch. 4.
The structural absence of genuine objection. Counsel only works if you can be told you are wrong by something immune to the urge to please you — so you have to command it to disagree, push back, and hunt for fatal flaws.
The False Finish (the half-second quit)
CoinedCoined by Daniel VanCott for the YAATZ framework, 2026.
In the narrative, Ch. 4 (the oxen) and Ch. 16.
The AI declares the work done while real work remains — wraps up a draft that's two-thirds there and calls it complete. Like the ox that quits a half-second early.
Operator Atrophy
CoinedCoined by Daniel VanCott for the YAATZ framework, July 2026.
Website only.
Your own skill wasting away from disuse as the AI carries the load. You don't notice, because the work keeps coming out good, so you lean on a muscle that has quietly hollowed out. The method is built to produce the opposite: the operator who stays in the chair comes out more capable.
Automation bias
BorrowedBorrowed (pre-LLM research term).
Website only.
Relying on the machine instead of checking it — either missing what it didn't flag, or following a wrong recommendation because it was the recommendation.