TL;DR
Most AI learning prompts help you recognize ideas. This one trains recall — via a paradox-first story, one lethal number, one mnemonic, and a three-stage interrogation (kid → pragmatic auntie → devil's advocate).
No app. No API. Two Markdown files. Copy, paste, learn.
The problem with "summarize this for me"
You ask ChatGPT to explain inflation. It gives a clean definition. You nod. You close the tab. Two weeks later — blank.
Recognition ≠ recall. Highlighting, mind maps, and AI summaries optimize for the wrong muscle.
Motif Learning Protocol v3.1 is my attempt to fix that with structured prompts — the kind of thing that belongs on dev.to because the real innovation is prompt architecture, not another flashcard app.
Core idea: find the paradox, not the definition
A motif here means a survival paradox — something that feels physically wrong but is true:
More money → less bread you can buy. (Inflation)
Brains ignore abstract definitions. They latch onto contradictions. The protocol forces every concept through that filter before anything else.
The four-step loop (Motif Tutor role)
| Step | What | Inflation example |
|---|---|---|
| Teach | Life fable with paradox baked in | King prints gold; bakers raise prices |
| Distill | 1 lethal number + 1 line mnemonic |
80%; "more money = less bread" |
| Test | Progressive pressure (3 personas) | "Your salary rose and eggs got pricier — is that inflation?" |
| Bind (optional) | Attach mnemonic to a daily physical action | Mumble the line when you open your wallet |
Step 3 is the differentiator. Not "do you understand?" but:
- 5-year-old — explain the motif in your own words, 2 sentences max
- Market auntie — boundary cases: applies / doesn't / partially
- Devil's advocate — counterexample: "Japan printed money — why no hyperinflation?"
Fail any stage → error attribution (what you confused), roll back to Teach. No participation trophies.
v3's secret weapon: pre-output gate in a code block
Most learning prompts list rules in prose. Models skim and ignore them.
This protocol requires the model to run a visible checklist inside a code block before every reply:
[思考过程]
1. What role am I? Which flow?
2. For this input: do what first, then what, then output what?
3. What's my output structure?
4. Role-specific checks — did I pass them?
That's a pre-compile check for pure prompts. Math Coach adds "did I give the answer?" Feynman Diagnostician adds "did I supplement knowledge instead of only asking?"
v2 → v3 reliability gains came mostly from this layer — not from adding more steps.
Five roles, one core prompt
| Role | Job |
|---|---|
| Motif Tutor | Full 4-step loop |
| Math Coach | Socratic — questions only |
| Concept Unpacker | Life analogy, 5-year-old readable |
| Devil's Advocate | Attack from 3 angles |
| Feynman Diagnostician | Probe blind spots, zero teaching |
Line 1 of the core prompt picks the role. Slashes like /rewrite, /skip, /memory-card work mid-session.
Drift recovery is first-class: one-line corrective prompts when the model dumps everything at once, hallucinates a paradox, or Math Coach "helpfully" reveals the solution.
Two-tier docs (lite vs full)
| File | Lines | Use when |
|---|---|---|
learning-prompts-lite.md |
~90 | Daily driver |
learning-prompts.md |
~870 | Article ingestion, full step templates, inflation walkthrough |
Progressive disclosure — don't make users read 800 lines to learn one concept.
Honest scope limits
Works well: causal / threshold / counter-intuitive knowledge — economics, systems design, engineering tradeoffs.
Skip the 4-step loop: pure how-to (Git commands), news, names/dates, concepts with no honest paradox (split or pick an adjacent concept).
Article entry path includes dehydrate → triage: if it's actionable checklist material, stop there. Don't force a fable onto an Excel tutorial.
Try it in 60 seconds
- Open learning-prompts-lite.md
- Copy the core prompt block into Claude / ChatGPT / Cursor
- Say:
Use Motif Tutor to help me learn "marginal utility"
Repo (public): https://github.com/zxpmail/learn-skill
Why this belongs on dev.to
- Zero runtime — it's prompt engineering as the product
- Pre-output gates + drift recovery = patterns you can steal for other agents
- Role dispatch + shared core mechanisms = lightweight multi-agent without a framework
- The inflation appendix is a golden-output fixture — useful for eval/regression if you fork this
If you've built learning agents and hit the "model nods along then forgets everything" wall — star the repo or steal the checklist pattern. Issues and PRs welcome.
Top comments (0)