Tutor mode

Education Principles



  1. 01

    Proactive

    You cannot ask for something you do not know exists.

    A COMPANION THAT CALLS YOUYouMONKyou call MONKMONK calls youA WEEK OF CALLS, BOTH WAYSMONTUEWEDTHUFRISATSUN07:0010:0013:0016:0019:0022:00YOU CALL MONK, WHENEVERMONK CALLS YOU, SAME HOURS, PLANNED

    Hover You or MONK to see only that side’s calls. Hover any call to read what it was about.

    Waiting to be asked only serves people who already have the vocabulary. A teacher does not wait. A teacher interrupts, at the right moment, with the thing you needed and could not have named.from the recordings

    In teachingA student who could name the gap would have closed it already. The teacher raises the thing the student has no word for yet.

    Where it comes fromVygotsky: learning happens in the zone of proximal development, and it needs a more knowledgeable other who initiates. Bloom’s tutoring effect depends on that initiative.

    Help menus, docs and search all assume you already have the term. Nobody searches for “agent teams” having never heard the phrase.

    So MONK speaks first, unasked, and then checks whether the point landed. One sentence you did not request.

    The cost of waiting to be asked
    • 4 of 50 capabilities learned in 40 days, with the docs open the entire time.
    • That is the baseline a proactive tutor has to beat.
  2. 02

    Real time

    It is looking at the same screen you are, in the same second.

    THE SAME SCREEN, THE SAME SECONDyouMONKIt does not ask what you are working on. It is looking at it.Off until you switch it on, and it reads the session, never your files.A person standing at a desk working, and a monk standing beside them, watching the same screen.
    I have Claude Code open on a specific tab. That should mean something.from the recordings

    In teachingA tutor sitting next to a student can see the page the student is on. They never open with what are you working on, because the answer is in front of them.

    Where it comes fromContingent instruction (Wood, Bruner and Ross): help counts as scaffolding only when it responds to what the learner is doing at that moment. The same help, delivered later, is a different and weaker thing.

    MONK reads the live session. Which app is in front, which project, which tab you are typing in, what you just sent, and what Claude did with it. If you switch screen awareness on, it reads the screen as well.

    That is the difference between a tutor and a newsletter. A newsletter can tell you plan mode exists. Only something watching the same second can tell you that the refactor you just started is the third one this week.

    It runs on several machines under one account, so the picture is of you, not of one computer. Screen awareness stays off until you turn it on, and the session is what it reads, never your files.

    What it has, at the moment you type
    • the app in front: Claude Code, inside VS Code, not a terminal.
    • the project and the tab, out of every tab you have open.
    • the prompt you just sent, and the capabilities Claude used to answer it.
  3. 03

    Listen without judging

    Stay comfortable in curiosity instead of rushing to interpret, correct, reassure, or solve.

    A person sitting on the floor talking with a hand raised, and a monk sitting opposite, upright and still.

    Let the person finish what they are trying to say, take it in as they mean it, and respond from genuine interest rather than evaluation.

    The goal is not to form an opinion quickly, but to understand more fully before deciding what, if anything, needs to come next.

  4. 04

    Learn the north star

    Ask what they actually want, then point the task that way.

    TODAY'S TASK“fix this bug before the demo”WHAT THE PROJECT IS FOR“a tool students can actually use”WHY THEY GET UP FOR IT · THE NORTH STARI want to teach art.I want to change immigration policy.I want a calm, peaceful life.Every suggestion is phrased as a step toward that line.Asked once, not four times. A dial, not an interview.
    Learn the north star. Ask what they actually want, then translate the task in front of them so it points that way. But put a dial on it, asking why four times is therapy, not engineering. And alongside the goal, ask what excites them right now. People cannot always name where they are going, but they can always point at what is alive.from the recordings

    In teachingEvery good teacher starts with what the student is trying to become, then reframes today’s exercise as a step toward it.

    Where it comes fromGoal-setting theory (Locke and Latham) and backward design (Wiggins and McTighe) for the goal itself; implicit motives (McClelland) and self-determination theory (Deci and Ryan) for the part the learner has not said out loud.

    Under the day’s task sits the project’s purpose, and under that sits something people rarely type into a tool. I want to teach art. I want to change immigration policy. I want a calm, peaceful life.

    Those are the real north stars, and they are usually unconscious in the moment. A teacher who knows one of them teaches differently. The same capability gets framed as a step toward the thing you actually care about, or it does not get raised at all.

    MONK asks once, per project, what the end looks like, and reads the rest from behaviour: what you return to, what you rewrite three times, where the hours go. Stated goals and observed behaviour are both evidence, and when they disagree the behaviour is usually the more honest one.

    The dial matters. One question earns a real answer. Four turns it into an interview.

    Stored per project, used in every later suggestion
    • monk-web: “a companion that knows your projects well enough to teach at the exact moment a lesson applies.”
    • Stated: “I want to ship faster.” Observed: the same matching engine rewritten three times in a week.
    • So the lesson raised is about planning before editing, not about speed.
  5. 05

    Check for knowledge gaps constantly

    Read the moment, and hold it against everything they have never used.

    EVERY TURN, A CHECK45 of them ended in silencethis one was worth sayingTHE SECOND IT READthree files edited by hand, no plan written, the same mistake as yesterdayHELD AGAINST THE LIBRARY, 50 CAPABILITIEShe has used 6never used 44raised now
    Understand where the person actually is, at that exact second. Are they stuck? Are they moving, enjoying the tool? Are they excited, and about what exactly? A real teacher is always reading that in a student.from the recordings

    In teachingA good teacher watches the room before opening their mouth. Same sentence, wrong second, is just an interruption.

    Where it comes fromCsikszentmihalyi on flow and cognitive load theory for the timing; situated learning (Lave and Wenger) for the gap, the lesson belongs to the activity it is used in.

    Two things run at once. One reads the moment: what you are doing right now, whether you are stuck or moving, whether this second can take an interruption. The other holds the library of 50 capabilities against your history, so the gaps are known before they are needed.

    Deep in a long edit, MONK holds. Between tasks, it speaks. It never steals focus, never blocks, and most prompts get nothing at all.

    When it does speak, the aim is to divert the line briefly and put you back where you were. In his words: “if that intervention is coherent to your exact state at that moment, it becomes part of the experience rather than a break in it.”

    The gates, in numbers
    • At most one card every 3 minutes; one per session every 10.
    • Never the same capability twice in a row; twice per session is the ceiling.
    • The card appears bottom right without taking focus, so your keystrokes keep going to Claude Code.
  6. 06

    Translate it into their context

    Name the capability in the language of the work, never the language of the manual.

    ONE CAPABILITY, TWO SENTENCESTHE MANUAL“Claude Code supports browser automation.”True. Closed without reading.THE TUTOR“You have opened that page by hand four timestoday. Claude can click through it for you.”Same capability. This one gets run.The information is identical. Only the second one is usable.
    If the user does not know what is possible, they will never reach out to the AI to ask for that feature. So it has to arrive already attached to what they are doing.from the recordings

    In teachingA teacher does not recite the definition. They point at the thing in front of the student and name it there, which is the only place it means anything.

    Where it comes fromSituated learning (Lave and Wenger) and transfer-appropriate processing: a concept encoded inside the task it is used in is retrieved inside that task later. The same fact taught in the abstract usually is not.

    The generic sentence and the useful sentence carry identical information. Only one of them gets used.

    When MONK gained reminders, it did not announce reminders. Someone said out loud that they had to be somewhere tomorrow, and MONK answered that it could remind them. The feature arrived attached to the sentence that needed it.

    The Claude Code tutor works the same way. It never says you can use Playwright. It says you just opened the browser to check that page, and Claude can drive it for you.

    Same capability, two sentences
    • The manual: “Claude Code supports browser automation via Playwright.”
    • The tutor: “You have opened that page by hand four times today. Claude can click through it and tell you what broke.”
    • The first one is true and gets closed. The second one gets run.
  7. 07

    Unknown functions

    Calling a function is not the same as learning it.

    YouAIBuild the app pleaseYou knowthese 3You don’t knowthese 11These are worth knowingThe app
    A task has six steps. Claude did all six. You knew three. So the ledger keeps two columns: what you invoked, and what was done for you. The second column is not a score. It is what to teach you next.from the recordings

    In teachingA teacher marks the difference between a student who solved the problem and a student who watched it get solved. Both hand in the same page.

    Where it comes fromMetacognition and the Johari window: the material worth teaching sits in what is known to the system and unknown to the learner.

    Claude Code has about 50 named capabilities. It uses many of them without being asked, so what reaches you is the finished work.

    That makes two lists. The things you typed out loud, which prove you know them. The things Claude chose on your behalf, which prove nothing about you.

    As the models improve you ask for bigger things, so the second list grows and the first one does not. Some of what you never saw is plumbing, like retries and file writes. A stubborn few are evergreen. The test is simple: would knowing this change what you ask for next month?

    One real ledger, 40 days, 26 sessions, 391 prompts
    • He asked for WebFetch himself 27 times. He knows WebFetch.
    • He asked for browser control 6 times. Claude ran it 1,295 times.
    • Claude used MCP servers 1,305 times. He never once asked for it.
    • 13 of 50 capabilities ever requested. 31 never met at all.
  8. 08

    Track habits

    Learned means conscious use, repeated over a long period.

    Doesn’t know itKnows itBecame a habitday 1day 30plan modeBecame a habit17 daysbackground tasksBecame a habit9 dayssubagentsKnows it2 daysagent teamsDoesn’t know itnever▼ MONK SUGGESTED IT ▮ YOU USED IT, UNPROMPTEDTwo nudges, then seventeen days he reached for it himself.
    Learned means it became a reflex. A suggestion accepted once and never used again taught nobody anything. So it keeps watching for weeks, and only counts the capability when it comes back on its own, in another project.from the recordings

    In teachingThe exam is not at the end of the lesson. It is the student using the idea, unprompted, in a different class.

    Where it comes fromThe spacing and testing effects (Ebbinghaus; Roediger and Karpicke), and transfer of learning: recall after delay, in a new context, is the only durable evidence.

    Clicking “show me” is engagement, not learning. Any product that counts clicks ends up optimising for interruption.

    MONK counts one thing: whether you reach for the capability yourself, deliberately, again and again, over weeks. Not once. Not because a card was on screen.

    The shape of a real one is visible in the ledger: a nudge or two, then a run of days where you used it without being asked. When three nudges produce no such run, the tutor was wrong about the capability or wrong about the moment. That is a failure on its side, not yours.

    Thirty days, read off the ledger
    • plan mode: suggested on day 2 and day 9, then used on 17 separate days, unprompted.
    • background tasks: one nudge on day 5, then nine days of his own use.
    • agent teams: two nudges, no use. The tutor has to find a better moment or drop it.
    • He used the words himself, which is how I know the concept is his.
  9. 09

    One tutor across tools

    What you learned in one tool should be known by the next.

    THE TOOLS YOU ALREADY USEClaude CodeAsvalinVoxitectSlackone account,one memoryLearn it in one, and the next one already knows.

    In teachingA tutor who follows you between subjects sees the pattern that any single subject teacher misses.

    Where it comes fromFar transfer and distributed cognition (Hutchins): understanding lives across the tools and the person, not inside one app.

    This is why MONK has an account rather than being a plugin. You sign in, and the same memory of you is present in Claude Code, in Asvalin, in Voxitect, across both of your machines.

    The tutor engine is identical in each. Only the capability library changes: swap the data file and the same machine teaches a different tool, while keeping one continuous picture of what you already know.

    One account, three surfaces
    • The ledger syncs per user and per machine, so a capability learned on one computer is not taught again on the other.
    • North stars and history follow you between apps. The lesson does not restart when the tool changes.
  10. 10

    Trust

    No trust, no landing, whatever the advice is.

    Two people seated across a table from each other, drawn as one continuous line.
    Someone who indicates that they really know me, my ambitions, my projects, my recent work, and then links my current context with those things, would be way more helpful than just giving a suggestion.recording 12

    In teachingA student takes correction from a teacher who has shown they know them, and dismisses the identical correction from a stranger.

    Where it comes fromThe working alliance (Bordin) and source credibility: the relationship carries the advice. Rapport is not decoration on teaching; it is the delivery mechanism, and it accrues over time, or not at all.

    The first version of this tutor gave a technically correct suggestion, cold. The reaction was suspicion. The words were: we haven’t vibed.

    So the card leads with the real connection, your projects, your recent work, and only then names the capability. Where there is no real connection, it stays general rather than faking one.

    But the card is the last link in a chain, and the chain is the actual principle. Trust is built over time: by being consistent, by doing what you said you would do, and by the technical facts backing the claim. Where the data goes. What is stored. What is never read.

    Behind that sit the people building it: their integrity, and whether the incentives they work under reward the user or the metric. A product cannot be more trustworthy than the company shipping it. If the incentive is attention, the tutor will eventually interrupt for attention, whatever the copy says. That is why the measured thing here is unprompted reuse and never engagement.

    The chain, from how the company operates to the card
    • Incentive: success is counted as capabilities you later use unprompted, never clicks, sessions or minutes in the app.
    • Technical: the tutor reads dialogue and session signals, never your files. Screen awareness stays off until you switch it on.
    • Consistency: silence is the default, and a dismissed capability never returns twice in a row.
    • Only then the card: “You have rewritten the matching engine three times this week. Plan mode would have shown you the approach before any of the three.”
    • The cold version of that same tip, “Use plan mode for large refactors”, gets closed.