Learning Accelerator

Learning Accelerator

Overview

Converts "help me learn X" into a structured system instead of a wall of text. Six modes — PLAN, LADDER, RESOURCES, CHEATSHEET, QUIZ, FEYNMAN — used alone on request or chained as a full loop. Every mode enforces retrieval over re-reading (TRAP-aligned: Test / Retain / Associate / Perform) and logs weak areas for spaced retest.

Usage

Pick the mode from what the user asks. When they just say "teach me X" or "I need to learn X", run the FULL LOOP: PLAN → LADDER → RESOURCES, then per session CHEATSHEET → QUIZ → FEYNMAN.

Mode triggers

• "learn X quickly / from scratch / in N hours" → PLAN
• "roadmap / path / where do I start" → LADDER
• "what should I read / watch" → RESOURCES
• "summarize / cheat sheet / revision" → CHEATSHEET
• "test me / quiz me / am I ready" → QUIZ
• "do I really understand X" / after any QUIZ round → FEYNMAN

Core Concepts

1. PLAN — 20-hour 80/20 plan

Identify the ~20% of concepts that deliver ~80% of practical results; be explicit about what is EXCLUDED and why. Break into 10 sessions × 2h (scale the count to the user's stated time budget), each session = objective + concrete deliverable + the last 15 minutes reserved for closed-book review of everything so far. Name the best resource per session.

2. LADDER — 5 difficulty levels

Break the topic into 5 levels, beginner → advanced. Each level: milestone (observable, testable — "can implement LRU from memory", not "understands caching"), core skills, and an exit test. The user MUST pass the exit test before advancing — no calendar-based promotion.

3. RESOURCES — top 5, justified

Top 5 books/videos/courses/sites for the topic, each with WHY it earns its time slot and what to skip inside it. Prefer FREE resources; for DSA/interview prep default to takeuforward.org, labuladong.online/algo, and neetcode.io. Verify links exist before recommending; NEVER invent titles or URLs.

4. CHEATSHEET — one page, 5-minute review

Single page: bullets, one ASCII/mermaid diagram where structure matters, and one practical example per concept. Hard cap: reviewable in under 5 minutes. For recurring topics, save it as a file (or artifact) so it persists — a cheat sheet that scrolls away is not a cheat sheet.

5. QUIZ — escalate until it breaks

10 questions, strictly increasing difficulty, ONE AT A TIME — never dump all 10. After each answer: grade it (correct / partial / wrong), explain the mistake mechanism, then adapt: wrong → next question probes the same weakness from a new angle; correct → escalate. Log every miss (date, topic, mistake, improvement, retest-due) to the user's error log. If the user has configured an error-log path (for example prep/error-log.md in their workspace), append there and honour its retest rule (default: retest in +3 days; a pattern exits the log after 2 consecutive clean closed-book solves).

6. FEYNMAN — teach it back

Explain the concept in the simplest accurate form (no jargon unless defined in-line). Then have the user explain it BACK; diagnose gaps in their version (missing mechanism, wrong causality, memorised phrase without a model); reteach ONLY the weak points; repeat until their explanation stands alone. The user has not learned it until they can teach it.

Rules

• Retrieval beats re-reading: every session ends with closed-book recall, not a summary re-read.
• Associate: tie each new concept to something the user already knows or has built (one sentence).
• Never mark a level or session complete on time spent — only on a passed exit test.
• Interview-prep topics inherit stricter rules automatically: time-boxed Perform conditions matching real loops, same-day error logging, free resources first.
• Never fabricate resources, links, or "commonly asked" questions. If a resource cannot be verified, say so and offer a verified alternative.

Example

User: "I need to learn Kafka quickly."

→ PLAN: 10 × 2h sessions on the 20% core (log/partition model, producer acks, consumer groups + rebalancing, offsets/exactly-once, retention/compaction); EXCLUDED: Streams DSL internals, Connect plugins (state why). Session 1 deliverable: local single-broker + CLI produce/consume; last 15 min closed-book: draw the partition/offset model from memory.

→ QUIZ q1: "What does acks=all guarantee, and what can still be lost?" … escalating to q10: "Design the consumer-group strategy for strict per-key ordering at 100K msg/s with rolling deploys."

→ FEYNMAN: "Explain why a consumer group can never have more active consumers than partitions — as if to a new grad."