The Card Describes A Mode Claude Code Ships. It Is One Command Away.
Most of the cards we go through are about features. This one is about teaching, which is rarer and harder to do well: seven steps for turning “I want to do X but I don’t know how” into something you can repeat next week without help. The pedagogy is good. Step 2 in particular — ask for the whole roadmap, then walk one step — is the kind of advice that works on people, not just prompts.
But read the verbs. Five of the seven steps describe something you do. Three bullets describe something Claude does:
- Step 3: “Claude will not move forward until you finish the current step.”
- Step 4: “Even when Claude could finish the task in seconds, it holds back.”
- Step 5: “When a step is done, say ‘done’ or ‘finished.’ Claude waits for this.”
And then, one line below that last one, the card says this:
“If Claude moves ahead without your signal, pull it back to one step.”
Those two bullets are in the same box and they disagree. The second one is the true one — and the gap between them is the whole subject of this piece, because the behaviour the card is describing is real, it has a name, and it is one command away.
Source card: “How To Get Claude To Teach You Step By Step (7 Steps)” — BANI Academy, 20 October 2026. Name the goal, ask for the roadmap, start with step one, do the work yourself, confirm before moving on, ask for a repeat, build for independence.
View the original post on Facebook
Diagram: who is doing what in each bullet
| The bullet | Whose behaviour | Is it default? | What actually produces it |
|---|---|---|---|
| “Tell Claude exactly what outcome you want” (1) | Yours | — | You type it. Nothing to arrange |
| “Ask Claude to list every step from start to finish” (2) | Yours | — | You ask. Models are good at this |
| “Claude will not move forward until you finish the current step” (3) | Claude’s | No | An instruction you give — or a mode you select |
| “Even when Claude could finish the task in seconds, it holds back” (4) | Claude’s | No. The default is to help as much as it can | An instruction you give — or a mode you select |
| “Claude waits for this” (5) | Claude’s | No | Same |
| “Ask Claude to rephrase, slow down” (6) | Yours | — | You ask |
| “Do it again next week on your own” (7) | Yours | — | Nothing to arrange. This is the point of the whole card |
The mode has a name
Claude Code ships five output styles — Default, Proactive, Concise, Explanatory and Learning. The last one is the card, implemented. From the documentation, quoted exactly:
“In the Learning style, Claude adds the same
Insightblocks as the Explanatory style and also asks you to write some of the code. Claude handles routine implementation itself. When it reaches a piece with a real design decision, such as error handling, a data structure, or business logic with more than one valid approach, it leaves a few lines for you.”“Claude marks the spot with a
TODO(human)comment in the file, then sends a request that says what’s already built, what to write, and what to weigh… Claude then stops and waits. Write your code at theTODO(human)comment and tell Claude when you’re done. Claude responds with oneInsightabout your code and continues the task.”
Line that up against the card. Step 4 is “asks you to write some of the code… leaves a few lines for you”. Step 5 is “stops and waits… tell Claude when you’re done”. Step 3 is the same stop. The card did not invent this behaviour; it described a shipped mode without knowing it was one.
You turn it on with /output-style learning, or from /config, or by setting "outputStyle": "Learning" in a settings file. One detail worth catching: in a settings file the value is case-sensitive, and a near-miss such as learning silently gives you Default. The slash command does not care about case.
And the documentation is honest about what a style is not
Here is the sentence that explains the card’s own contradiction:
“An output style gives Claude instructions to follow. It doesn’t guarantee that something always happens or never happens.”
“It’s an instruction Claude follows, so nothing enforces it.”
So even with the mode on, “Claude waits for this” is a strong tendency rather than a rule — which is exactly why the card needed that second bullet about pulling it back. The card observed the right thing. It just filed it as a tip instead of as the limit of the mechanism.
The documentation then does something cards almost never do: it tells you what to use when an instruction is not enough.
| What you want | Use | Why |
|---|---|---|
| Restraint in this one conversation | Write it into the prompt — what the card is doing | Free, instant, and it works often enough to be worth doing. It is also forgotten the moment you open a new chat |
| Restraint for a whole session, every response | An output style — /output-style learning |
“It applies to the whole session, and you switch styles with one command”. Set outputStyle in ~/.claude/settings.json to make it your default everywhere |
| Something that has to happen every time, without exception | A hook | “Claude Code runs a hook itself at a lifecycle event, so it doesn’t depend on Claude following an instruction” |
| Claude to know your project’s conventions | CLAUDE.md | It stays loaded whichever style you pick |
Why the default pushes the other way
It is worth saying why restraint has to be requested, because it is not a flaw. A model asked “I want to build Y” is trying to be useful, and the useful thing — measured by almost any yardstick — is to build Y. Holding back so that a human learns is a different objective, and nothing in the request states it.
The prompting documentation shows the same pattern from the opposite side, describing models that “overengineer by creating extra files, adding unnecessary abstractions, or building in flexibility that wasn’t requested”. Its recommended fix is not to wait for better defaults. It is: “add specific guidance”. Say what you want less of.
That is the same move the card needs and does not make. “Claude holds back” is the outcome; “do not write this for me — tell me the one move and wait” is the instruction that produces it.
The seven steps, graded
| Step | What kind of advice it is | Verdict |
|---|---|---|
| 1. Name your goal clearly | Pedagogy. Outcome, not method; plain language | Sound. And “not the method, the result” is the right instinct — you do not yet know the method |
| 2. Ask for the full roadmap first | Pedagogy. See the map, walk one step | The best step on the card. It also gives you something to check the model against later |
| 3. Start with step one only | Half pedagogy, half behaviour claim | The advice holds; the promise does not. “Claude will not move forward” needs the mode or the instruction |
| 4. You do the work, not Claude | Behaviour claim | This is the Learning style, by name. It will not happen on its own |
| 5. Confirm before moving on | Behaviour claim, and it knows it | The second bullet quietly admits the first is unreliable. Both are true in the right order |
| 6. Ask to repeat if anything is unclear | Pedagogy. “There is no penalty for asking again” | Sound, and the thing most people need permission to hear |
| 7. Build for independence | Pedagogy. Do it again next week from memory | The point of the exercise, and the only measure of whether any of it worked |
The one-command version
Keep all seven steps. Add one line at the front and, if you are in Claude Code, one command.
/output-style learningAnd in a plain chat, where there is no such command, type the mode instead:
“Teach me to do this rather than doing it for me. Give me the full roadmap first, then one step at a time. For each step: tell me the exact move, then stop and wait for me to say ‘done’. Do not do the step for me even if it would take you seconds. If I say I do not understand, explain it a different way before moving on.”
That paragraph is the card, turned from a description of Claude into an instruction to Claude. Which is the one edit it needed.
Four honest caveats
- The Learning style is a Claude Code feature, and it is written around code. Its examples are
TODO(human)comments andvalidateFile(). The card is general — “automate X”, “build Y” — so in a plain chat you want the paragraph above rather than the command. - We did not run either version. This is the card read against the output-styles documentation, in October 2026.
- “Not default” is a claim about tendency, not capability. Ask any current model to hold back and it will try. The point is that it will not try unless asked, and the card tells you it will.
- This is a good card. Four of the seven steps are teaching advice that needs no machinery at all, and step 2 is better than most of what gets written about learning from a model. The fix is three bullets, not seven steps.
The one-line version
Three of the card’s bullets describe Claude holding back, waiting and refusing to move on — that is a named mode, /output-style learning, not a default; and even switched on it is an instruction rather than a guarantee, which is precisely why the card’s own step 5 has to tell you what to do when it does not hold.
Turn the description into an instruction. Then the other four steps do the teaching, which was always the good part.
Source card: “How To Get Claude To Teach You Step By Step” — BANI Academy, 20 October 2026.
View the original post on Facebook · Related: which typed words the software actually recognises and what loads from your .claude folder
