df2tm

what it does

it teaches while claude works. when a load-bearing concept shows up in your task, df2tm drops a short, skimmable aside explaining it, and then gets out of the way.

it remembers across sessions. concepts worth keeping go into a local learner model with spaced-repetition review dates, and next time they come up it checks what stuck.

it makes you a better director. beyond the code itself it teaches you to steer claude with sharper prompts: when to delegate, when to learn, when to ask for the reasoning.

the relevance gate

the hard part of a teaching layer is not teaching. it is silence.

every potential teaching moment passes a four-part gate before anything appears: load-bearing, forward-useful, novel-enough, and budget allows. pass it and you get an aside, skimmable at zero cost to the work. fail it, or say you are firefighting, and df2tm says nothing at all.

that gate is the product. an assistant that explains everything is one you turn off in a week.

memory

a session-start hook runs at the beginning of every session. it reads your learner model, surfaces the concepts due for review, and quietly activates.

concepts are recorded with a grasp level, new to shaky to solid, and a spaced-repetition due date, so later sessions reinforce what matters in that project or globally rather than repeating what already landed.

steering

the controls are plain language, not syntax. "teach me more" and "teach me less" move the intensity. "just do it" goes silent and stays silent. "why did you do that" explains the last decision, "quiz me" runs active recall on what is due, "debrief" recaps the session, and "i already know this" retires a concept for good.

there are slash commands under /df2tm: for the same things, but nothing requires them.

it also auto-calibrates. engage and it teaches a little more, ignore it and it backs off. the changes are always explained, never silent.

the science

the teaching draws on about 170 learning-science principles: evidence-based techniques such as spaced repetition, active recall, dual coding and elaborative interrogation, the cognitive effects that justify them, and the mechanisms underneath.

they are applied dynamically, based on you and on the work, never all at once. the repository lists the full library.

state and privacy

everything is local on your machine. nothing is sent anywhere.

two files hold it all: a learner model with your preferences and one line per tracked concept, and a dated journal of what you have been taught, per project. the state directory can be relocated to a synced folder, and deleting it resets the history completely, re-seeding on the next session.

the repository

it lives on github, opens in a new tab, with the install steps and the full library of principles.