Deliberate Practice

Overview

Most people confuse repetition with learning — they accumulate years of experience and plateau. Once an activity becomes automatic, executing it no longer builds new neural architecture. Ericsson, Krampe & Tesch-Römer (1993) showed the predictive variable is not hours of doing but hours of specifically deliberate practice — targeted, uncomfortable, feedback-rich repetition designed to build mental representations.

Cross-skill composition: Use feedback-loops first (audit your error signal); then metacognition (surface your current representation gap); use instead of deep-work when acquiring skills, not producing output; use alongside cognitive-evolution-stages for stage-aware practice design.

When to Use

Trigger: plateau despite experience; designing high-performance learning program; training hours high but skill transfer low; evaluating whether practice is building capability or maintaining it; skill atrophy or deskilling as AI copilots absorb the routine reps (AI adoption, AI hype, "will AI make me worse at my craft"). When NOT: goal is execution not acquisition (use deep-work); no expert benchmark exists; bottleneck is motivational not representational.

Coaching Novices (Adaptive Front Door)

Engine mode: user has a concrete case → run The Process directly. Coach mode: user is unfamiliar or has no concrete case → guide step by step.

In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.

1. Ask the plateau question: "When you practice X, what does it feel like after 15 minutes — harder, the same, or easier?" Comfort/easy = automatic = not building representations.
2. Find the expert performance structure: "Who is world-class at X? What do they perceive in the first 3 seconds that you don't?" This locates the mental representation gap.
3. Identify the discomfort zone: "What part of practicing X makes you most want to stop?" That is almost always where the gap lives.

[WAIT — do not advance until user responds]

4. Design the smallest feedback loop: "How would you know within 60 seconds whether a move was correct?" Latency over 24h kills representation-building.

[WAIT — do not advance until user responds]

5. Set the repetition target and stop-rule: "How many reps of this specific discomfort can you sustain before concentration drops?" (1–4 hours/day is Ericsson's ceiling.)

[WAIT — do not advance until user responds]

The Process

Step 1 — Define the sub-skill with precision. Not "get better at X" — specify the exact representational gap (e.g., "detect when counterpart shifts from positional to interest-based"). Step 2 — Find or construct the feedback mechanism. Latency >24h breaks action-result association. Expert feedback > peer one level above > simulation with ground truth. Step 3 — Diagnose the mental representation gap. Ask: "What does an expert see here that I don't?" Not what they do — the doing follows from the seeing. Step 4 — Design the repetition targeting the gap. Must trigger the sub-skill, produce in-session feedback, and be executable at dozens–hundreds of reps per session. Step 5 — Track representation progress, not output. Output metrics lag by weeks. Track: "Am I perceiving X earlier than before?" Step 6 — Apply the stop-rule. Comfort = automaticity maintenance. Redesign to a harder sub-skill. End session when concentration drops.

Output: Practice Design Artifact

Target sub-skill (precise): [specific representational gap]
Expert mental representation: [what expert perceives that I currently don't]
Current representation gap: [specific failure mode]
Feedback — Source / Latency / Reliability:
Repetition — Exercise / Volume / Duration / Frequency:
Progress indicator (representation-level, not output): [what I will perceive by Week N]
Stop-rule triggers: comfortable → redesign; concentration drops → end; latency >24h → redesign

→ Method in Action: Berlin Violin Study (1991–1993) · Franklin's Spectator Method → 2026 lens: Keeping Skill Alive When AI Does the Reps (2024–2026)

Practice Design Domain Packs

Medicine/Surgery: sub-skill: laparoscopic tissue manipulation; feedback: simulator + debrief within 1h; rationalization to reject: "I'll improve with more cases." Writing: sub-skill: eliminate nominalization in first-draft prose; feedback: rewrite published paragraphs vs original; rationalization: "I write every day." Investment: sub-skill: identify customer concentration risk from footnotes in 20 min; feedback: 50-case retrospective library with outcomes.

Contribute packs via the deciqAI repo — requires sub-skill, expert representation, feedback latency, and common rationalization.

Applying It Well

• Target representations, not outcomes — ask "What does the expert perceive that I don't?"
• Make feedback faster — redesign question: "How do I get a reliable signal within 60 seconds?"
• Comfort signals time to redesign, not celebrate.
• 1–4 genuine hours/day is Ericsson's hard ceiling; volume in degraded concentration reinforces errors.
• Require expert think-alouds or annotated examples — you cannot design practice you cannot see.

→ Primary sources: references/sources.md

Common Rationalizations

[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.

Red Flags / Verification

• Sessions feel comfortable — automaticity has absorbed the activity.
• Feedback latency measured in days — action-result association cannot form.
• Practitioner describes what expert does but not what they perceive — no representational target.
• Practice volume cited as expertise without verifying hours were deliberate.
• Sub-skill = specific representational gap; feedback latency <24h; expert representation identified.
• Reps: dozens per session; stop-rule applied; progress tracked at representation level not output level.

Part of deciqAI Knowledge Skills — 253 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/s/deliberate-practice · Built by deciqAI · github.com/deciqAI · Contributions welcome.

Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/deliberate-practice.json