2nd/10_Wiki/Topics/AI_and_ML/Deliberate-Practice.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-deliberate-practice	Deliberate Practice	10_Wiki/Topics	verified	self	deliberate practice Anders Ericsson expertise 10000 hours mental representation growth mindset	none	A	0.93	applied	learning deliberate-practice ericsson expertise mental-models skill-acquisition		2026-05-10	pending	language applicable_to learning theory Skill Acquisition Career Growth Expertise Development

Deliberate Practice

매 한 줄

"매 단순 반복 X — 매 한계 zone 의 precise 의 training". Anders Ericsson 의 "Peak" (2016). 매 specific goal + 매 feedback + 매 effort + 매 mental model. 매 10K hour 의 myth — 매 quality > quantity. 매 modern: 매 LLM-aided coach.

매 핵심 element

Ericsson 의 4 condition

1. Specific goal (not "improve").
2. Focused intensity (not background).
3. Immediate feedback (not delayed).
4. Comfort zone 의 escape (learning zone).

매 mental representation

• 매 expert 의 domain-specific structure.
• 매 chunk recognition (chess master).
• 매 deep pattern.

매 zone

• Comfort: 매 already mastered.
• Learning: 매 just-beyond — 매 deliberate target.
• Panic: 매 too far → 매 disengage.

매 vs general practice

측면	Deliberate	Casual
Goal	Specific	Vague
Effort	Maximum	Comfortable
Feedback	Immediate	None
Reflection	Yes	No
Mentor	Often	Rarely

매 limit

• 매 매 hour 4-5 max (cognitive fatigue).
• 매 mentor / coach 의 important.
• 매 self-practice 의 wrong technique 의 reinforce.
• 매 over-emphasis on natural talent ignored.

매 critique (recent research)

• 매 talent + practice 매 둘 다 important.
• 매 10K hour 의 average 가, 매 huge variance.
• 매 domain 의 따라 transfer 의 다름.

매 modern AI 의 응용

• Aim training (Cognitive Training Software (eg Aim Lab_KovaaKs)): 매 deliberate practice 의 gamified.
• Code review feedback: 매 immediate.
• LLM coach: 매 personalized feedback.
• Spaced repetition (Anki).
• Active recall.

매 plan

1. Identify weakness (specific).
2. Set measurable goal.
3. Design exercise (just-beyond).
4. Schedule (focused 4 hour max).
5. Get feedback (mentor / data / LLM).
6. Reflect (what worked / didn't).
7. Iterate.

💻 패턴

Practice plan template

skill: 'TypeScript advanced types'
current_level: 'intermediate'
target: 'use conditional types comfortably in 2 weeks'

weakness_audit:
  - 'don't recognize when to use conditional types'
  - 'syntax of `infer` 의 rough'
  - 'distributive conditional 의 confused'

weekly_schedule:
  - day: Mon
    focus: 'conditional type basic'
    exercise: 'solve 5 type challenges'
    duration_min: 60
    feedback: 'TypeScript Playground'
  - day: Wed
    focus: 'infer keyword'
    exercise: 'extract type from function signature'
    duration_min: 60
    feedback: 'review with senior'
  - day: Fri
    focus: 'apply to real codebase'
    exercise: 'refactor one util'
    duration_min: 90
    feedback: 'PR review'

reflection_after_each:
  - 'what was hardest?'
  - 'what mental model is forming?'
  - 'what is next constraint?'

LLM-aided coach

def deliberate_coach(user_attempt, reference, skill='code_review'):
    prompt = f"""You are a deliberate practice coach for {skill}.

User's attempt:
{user_attempt}

Reference (best practice):
{reference}

Provide:
1. Specific gap (1 sentence).
2. Root cause / missing mental model.
3. ONE concrete next exercise (just-beyond comfort).
4. What feedback signal to look for.

Don't praise. Be specific and constructive."""
    return llm.generate(prompt)

Spaced repetition (Anki-style)

class SpacedReview:
    def __init__(self):
        self.cards = {}  # 매 id → (next_review, interval, ease)
    
    def review(self, card_id, quality):  # 매 quality 0-5
        prev = self.cards.get(card_id, (datetime.now(), 1, 2.5))
        _, interval, ease = prev
        
        if quality < 3:
            interval = 1  # 매 reset
        else:
            ease = max(1.3, ease + 0.1 - (5 - quality) * 0.08)
            interval = int(interval * ease)
        
        self.cards[card_id] = (
            datetime.now() + timedelta(days=interval),
            interval, ease,
        )

Active recall (vs re-read)

def study_session(material, mode='active_recall'):
    if mode == 'active_recall':
        return [
            'Read chapter (no notes)',
            'Close book',
            'Write everything you remember (5 min)',
            'Compare with book — note gaps',
            'Re-read only the gap parts',
            'Test self again next day',
        ]
    elif mode == 'passive_reread':  # 매 ❌ less effective
        return ['Read 3 times']

Feedback loop (immediate)

class PracticeSession:
    def __init__(self, skill):
        self.skill = skill
        self.attempts = []
    
    def attempt(self, action):
        result = execute(action)
        feedback = evaluate(result, target=self.skill.target)
        self.attempts.append({
            'action': action,
            'result': result,
            'feedback': feedback,
            'timestamp': datetime.now(),
        })
        # 매 immediate 의 next attempt 의 inform.
        return feedback
    
    def reflect(self):
        return llm.generate(f"""Reflect on this practice session:
{self.attempts[-10:]}

What patterns? What's the next bottleneck?""")

Mental model construction

def chunk_practice(domain):
    """매 expert 의 chunking 의 build."""
    # 매 e.g., chess: 매 매 board configuration → 매 pattern recognize
    # 매 매 day 의 100 position → 매 quick recognize / response
    return [
        'Day 1-7: 매 100 pattern flashcard / day',
        'Day 8-14: 매 mixed deck (no order)',
        'Day 15+: 매 apply in game (transfer)',
    ]

Energy management

def deliberate_practice_schedule():
    """매 cognitive fatigue 의 limit."""
    return {
        'morning_block': '90 min hardest skill (peak focus)',
        'break': '15 min walk (DMN-friendly)',
        'mid_block': '90 min secondary skill',
        'long_break': '60 min lunch + nap',
        'afternoon_block': '60 min review + reflect',
        'rest_of_day': 'light tasks / non-deliberate',
        'total_deliberate': '~4 hours max',
        'days_off': '1-2 / week (recovery)',
    }

Skill progression (Bloom's Taxonomy + Dreyfus)

DREYFUS_STAGES = [
    ('Novice', 'Rule-following, hand-holding'),
    ('Advanced Beginner', 'Recognize aspects'),
    ('Competent', 'Plan + analyze, follow standard process'),
    ('Proficient', 'Holistic + experience-driven'),
    ('Expert', 'Intuitive + transcend rules'),
]

def practice_for_stage(stage):
    """매 매 stage 의 different practice."""
    return {
        'Novice': 'Tutorials + close mentorship',
        'Advanced Beginner': 'Guided practice + frequent feedback',
        'Competent': 'Independent project + post-mortem',
        'Proficient': 'Complex challenge + reflection',
        'Expert': 'Teaching + writing + research',
    }[stage]

매 결정 기준

상황	Approach
New skill	Tutorial + mentor + small wins
Plateau	Identify specific weakness + targeted exercise
Career growth	T-shape: 매 deep + 매 broad
Esports	Aim trainer + mentor (review video)
Code	TDD + code review + refactor practice
Writing	Daily output + edit + critic
Music	Slow + isolated + correct rep

기본값: 매 specific + 매 feedback + 매 4-hour cap + 매 reflection.

🔗 Graph

• 부모: Expertise
• 응용: Cognitive Training Software (eg Aim Lab_KovaaKs) · Pair-Programming
• Adjacent: Default Mode Network (DMN) · Brain-Derived Neurotrophic Factor (BDNF) · Cognitive Reserve Theory · Articulateness · Bounded_Rationality
• 사상가: Anders-Ericsson · Carol-Dweck (growth mindset)

🤖 LLM 활용

언제: 매 skill development plan. 매 plateau breaking. 매 feedback design. 매 personal coach. 언제 X: 매 fixed mindset 의 reinforce. 매 burnout 의 over-push.

❌ 안티패턴

• Mindless rep: 매 habit 의 X.
• Comfort zone 의 stuck: 매 plateau.
• No feedback: 매 wrong technique 의 reinforce.
• Over-train (>4hr): 매 fatigue 의 quality drop.
• Talent 의 over-emphasize: 매 fixed mindset.
• Solo practice 의 only: 매 mentor 의 가치 lose.
• No reflection: 매 learning 의 X.

🧪 검증 / 중복

• Verified (Ericsson "Peak", Dreyfus stages, Bloom taxonomy).
• 신뢰도 A.
• Related: Cognitive Training Software (eg Aim Lab_KovaaKs) · Default Mode Network (DMN) · Brain-Derived Neurotrophic Factor (BDNF) · Cognitive Reserve Theory · Articulateness · Be-Detailed.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — Ericsson 4 + 매 zone + plan + LLM coach + Dreyfus + spaced rep code