Every interface shift removed friction — and a reason to think. Tuukul trains the judgment that remains: what you hand to AI, what you accept back, and what you sign your name to.
Punch cards
thinking was mandatory — every run cost a day
Terminals
syntax was the gatekeeper
GUIs & IDEs
the machine started meeting you halfway
Autocomplete
the machine started finishing your sentences
Chat & agents
you describe outcomes; software does the work
What's next
when the interface disappears, thinking is all that's left
Working with AI is a chain of judgment calls. Tuukul installs a deliberate pause at each of the three moments that decide everything.
Delegate
before the AI runs
“What exactly am I asking for — and what would tell me the answer is wrong?”
Frame the outcome, state a prediction with a confidence, bound the blast radius.
Accept
when output returns
“Did I verify this, or did it just sound right?”
Check before adopting: run it, trace the claim, reconcile against your prediction.
Deploy
before the deed
“If this ships and fails, who is affected — and did I sign it?”
Premortem consequential actions, check reversibility, disclose honestly.
Before a consequential AI run, Tuukul asks for a one-line prediction and a confidence. Afterwards it reconciles what actually happened. Over weeks that becomes your personal calibration curve — computed from real work, not a quiz. Nobody else can show you this.
The free Lab uses general knowledge. The real thing measures your calls on your own AI work.
No abstract brain-training — research is clear that puzzle scores don't transfer. Every rep in Tuukul runs on the work you were already doing today.
Morning: what's the most consequential thing you'll delegate to AI today — and how confident are you? Evening: what actually happened? Reconcile, reflect, done in three minutes.
A Socratic coach that answers your statement with a question — why this task, why that confidence, what would change your mind. Never more than two probes before it gives you value.
Short drills on your real backlog: decomposition katas, premortems on your actual next launch, estimation ladders, inversion, war-room scenarios.
Auto-drafted from your sessions; you confirm in a minute. Over months it becomes the evidence portfolio behind your credential.
Calibration, gate discipline and verification-before-accepting feed the thinking strand of your one Itz'at Score — every number carries a “how computed” note.
The same coach rides along in Claude Code and your IDE — nudging at the gates of real runs, not in another tab you'll never open.
Tuukul's ingestion schema has no field for your prompts — a privacy promise you can verify, not just trust. You own your memory: view it, edit it, export it, erase it.
Miatz Tuukul wants to coach you
This is the real consent screen — not a marketing mock.
What it reads
What never leaves your machine
Use of AI is soaring; evaluation of AI is not. These are the numbers — each links to its source.
Join the waitlist and you'll get the launch note, the open Itz'at rubric, and first access as cohorts open.
Tuukul (Yucatec Maya: thought) is the structured-thinking strand of miatz's AI-Fluency program. It trains judgment practices — framing what you delegate, verifying what you accept, signing what you deploy — and measures calibration from your real AI work. It feeds one score (the Itz'at Score) and one credential (Itz'at AI-Fluency).
No. Raw prompt and response text never leaves your machine — the upload schema has no field for it. Only features (counts, flags, your own short predictions) sync, and you can pause or delete everything at any time.
Aggregates only — cohort-level averages over groups of five or more. Individual coaching stays between you and the coach unless you explicitly opt in to identified sharing.
Never. The coach asks at most two questions before giving value, and it never rewrites your prompt or hands you answers. Teach, don't do — that's the design law.
The Calibration Lab shows you your confidence-vs-accuracy gap right now — free, no account.