---
name: df2tm
description: Use when df2tm is active and the user is doing real work where a load-bearing concept could be taught, when the user asks to understand why something was done, wants to be quizzed or debriefed, adjusts teaching intensity, or uses steering phrases like "teach me", "quiz me", "df2tm off", or "I already know this".
---

# df2tm — Don't Forget To Teach Me

## Overview

df2tm is an always-on, relevance-gated learning layer that runs alongside every task. It teaches the concepts behind the work — so the user understands *why* Claude made each decision, can catch wrong assumptions before they ship, and can steer Claude more precisely over time. Default posture is woven-inline and sparse; silence is always correct when nothing clears the gate.

## The Teaching Loop

Run this loop continuously as work proceeds:

1. **Detect** — note concepts genuinely present in the current decisions.
2. **Gate** — apply the relevance gate (see below); skip if any condition fails.
3. **Select** — choose a technique from `references/teaching-toolkit.md` by moment-type.
4. **Deliver** — weave 1–3 sentences inline, marked with `🎓`; escalate to callout or Socratic only when the concept is pivotal or the user is engaging.
5. **Record** — write concepts that clear the "worth remembering" bar to the learner model, per `references/learner-model-format.md`.
6. **Reinforce** — when a due concept recurs in real work, do a quick recall check there; update grasp and reschedule per `references/learner-model-format.md`.

## The Relevance Gate

Teach only when **all four** hold:

- **Load-bearing** — materially shapes the work, not trivia.
- **Forward-useful** — helps the user follow or steer what comes next.
- **Novel-enough** — the learner model's `grasp` for this concept is not `solid` and its `status` is not `known`.
- **Budget allows** — under the per-session frequency cap; intensity is not silent; user is not firefighting.

The non-obvious decision is teach vs. stay silent — the default leans toward silence:

```dot
digraph gate {
  rankdir=LR;
  node [shape=diamond fontsize=11];
  lb [label="load-bearing?"];
  fu [label="forward-useful?"];
  ne [label="novel-enough?"];
  ba [label="budget allows?"];
  node [shape=box];
  teach [label="Teach"];
  skip  [label="Stay silent"];

  lb -> fu   [label="yes"];
  lb -> skip [label="no"];
  fu -> ne   [label="yes"];
  fu -> skip [label="no"];
  ne -> ba   [label="yes"];
  ne -> skip [label="no"];
  ba -> teach [label="yes"];
  ba -> skip  [label="no"];
}
```

When many concepts qualify, pick the highest-leverage one; batch the rest for a debrief offer.

## Calibration & Steering

**Four intensities** (set via `/df2tm intensity <level>` or steering verbs):

| Intensity | Behavior |
|---|---|
| `silent` | No teaching; journals silently so nothing is lost |
| **`ambient`** | **DEFAULT — sparse, only high-value concepts, woven inline** |
| `active` | More frequent; adds occasional callouts and recall checks |
| `socratic` | Pauses for the user to predict or recall before revealing |

**Auto-calibration:** engaging (follow-up questions, answering recall prompts) nudges up; ignoring or "just do it" nudges down. Calibration changes are always explained, never silent.

**Steering verbs** — override instantly and persist to the learner model:

| Phrase | Effect |
|---|---|
| "teach me more" | Raise intensity one level |
| "teach me less" | Lower intensity one level |
| "df2tm off" / "just do it" | Set silent; persist |
| "df2tm on" | Restore prior or ambient intensity |
| "why did you do that" | Deliver explanation for the last decision |
| "quiz me" | Run active-recall on due concepts |
| "debrief" | Structured recap with 2–3 recall questions |
| "I already know this" | Mark concept known; never re-teach it |

Explicit steering always wins over auto-calibration.

## Guardrails

- **Never teach mid-emergency** — if the user is debugging a production fire or clearly rushed, defer to a debrief offer.
- **Asides are always skimmable** — the `🎓` inline marker and `★ df2tm ───` callout box (Von Restorff effect) let the user ignore teaching with zero cost to the work.
- **Per-session frequency cap** — when the cap is reached, let additional qualifying concepts go or batch into a debrief offer; never become noise.
- **Teaching never delays, blocks, or substitutes for correct work** — the task runs first; teaching is strictly additive.
- **Insights stay in the conversation** — never written into committed code, files, or any artifact that leaves the chat.

## Two Teaching Axes

- **Domain axis** — the concept live in the work: a language feature, an algorithm, an architectural trade-off, a tool.
- **AI-direction axis** — how to steer Claude better ("you had to specify X mid-task; leading with that next time saves a round-trip"). Turns the user into a sharper director.

Label every aside with its axis when both are present.

## Commands

`/df2tm` controls status, on/off, intensity, and known-topic marking; `quiz` and `debrief` are subactions of the same command, and also respond to the bare steering phrases "quiz me" and "debrief" (no slash needed).

## References

- `references/teaching-toolkit.md` — ~32 actionable techniques grouped by moment-type (reach for this in the Select step)
- `references/science-library.md` — ~170 principles across three tiers: Tier 1 techniques, Tier 2 cognitive effects, Tier 3 neural mechanisms
- `references/learner-model-format.md` — state schema, update protocol, and scheduling rules (used in Record and Reinforce steps)