Pro Workflow — Battle-Tested Claude Code Patterns

01 Core Architecture

Command → Agent → Skill

Command

User input
Slash commands

→

Agent

Orchestrate
Plan & Review

→

Skill

Execute
Domain logic

Agents cannot call each other directly — use Task(subagent_type=...)
Progressive skill disclosure: only surface what’s needed now
Skills use context: fork for isolation

02 The 8 Patterns

Karpathy’s 80/20 Ratio

Pattern	What It Does
Self-Correction	Claude learns from corrections automatically
Parallel Worktrees	Zero dead time — native `claude -w`
Wrap-Up Ritual	End sessions with intention, capture learnings
Split Memory	Modular CLAUDE.md for complex projects
80/20 Review	Batch reviews at checkpoints, not every change
Model Selection	Opus 4.6 & Sonnet 4.6 adaptive thinking
Context Discipline	Manage your 200k token budget
Learning Log	Auto-document insights from sessions

03 Components

Everything Included

9

Skills

3

Agents

12

Hooks

10

Commands

3

Contexts

6

Rules

Platforms

Claude Code Cursor SkillKit (32+ agents)

Database

SQLite + FTS5 10 Categories Full-Text Search

04 Self-Correction Loop

Compounding Improvements

1

Correct

User fixes mistake

→

2

Propose

[LEARN] rule

→

3

Approve

User confirms

→

4

Persist

LEARNED.md / DB

“Small corrections compound into massive accuracy gains over time.”

Remember this Add to rules Don’t do that again

05 Split Memory

Modular CLAUDE.md

CLAUDE.md

Entry point — imports all modules

AGENTS.md

Workflow rules

SOUL.md

Style prefs

LEARNED.md

Auto-populated

Context Discipline

Rule	Why
Read before edit	Understand context
Compact at boundaries	Free token budget
<10 MCPs, <80 tools	Reduce noise
Subagents isolate	Protect main context
CLAUDE.md ≤ 150 lines	Avoid instruction drift

06 Hooks & Automation

12 Automated Hooks

PreToolUse

Edit tracking, quality gates

Pre-Commit

Lint & typecheck reminder

PostToolUse

console.log, secrets scan

Test Failure

Auto-learn from failures

SessionStart

Load learnings, show stats

SessionEnd

Save stats to SQLite

PromptSubmit

Drift detection at 6+ edits

PreCompact

Save state before compact

Stop

Context-aware reminders

ConfigChange

Detect settings drift

Notification

Permission logging

Pre

Post

Session

Other

07 Specialized Agents

3 Purpose-Built Agents

Planner

Read-Only

Break down complex tasks into step-by-step plans. Never edits — only reads and proposes.

Multi-file changes Architecture >10 tool calls

Reviewer

Quality Gate

Logic, edge cases, error handling, security, performance, and test coverage audit.

Critical High Medium Low

Scout

Confidence Gate

Assess readiness with worktree isolation. Scores 0–100 across 5 dimensions.

08 Scout: Confidence Scoring

5 Dimensions × 20 pts

Scope Clarity

20

Pattern Familiarity

20

Dependency Aware

20

Edge Case Coverage

20

Test Strategy

20

GO

≥70 — Proceed with confidence

HOLD

<70 — Identify gaps first

Context Modes

Dev

Code first
Iterate quickly

Review

Security focus
Read-only audit

Research

Explore broadly
Summarize findings

09 Development Workflow

Research → Plan → Implement

R

Research

Feasibility analysis. Scout scores readiness.
Output: GO / HOLD decision

P

Plan

Product requirements + architecture.
Planner agent breaks down steps

I

Implement

Phased execution with frequent commits.
Reviewer validates at checkpoints

Validation gate between each phase — must pass before proceeding.

“80% of my code is written by AI, 20% is spent reviewing and correcting it.”

— Andrej Karpathy

10 Skills Library

9 Battle-Tested Skills

Skill	What It Does
pro-workflow	Core 8 patterns
smart-commit	Quality gates + conventional commits
wrap-up	Session ritual with learning capture
learn-rule	Persist corrections to memory
parallel-worktrees	Git worktree setup for zero dead time
replay-learnings	Surface relevant past patterns
session-handoff	Resume docs for next session
insights	Analytics, heatmaps, trends
deslop	Remove AI code slop before commit

11 Model Selection

Right Model, Right Task

Task	Model
Quick fixes	Haiku 4.5
Features	Sonnet 4.6 (adaptive thinking)
Refactors	Opus 4.6 (adaptive thinking)
Architecture	Opus 4.6 (1M context beta)
Hard bugs	Opus 4.6 (1M context beta)

Wrap-Up Ritual (5 Steps)

1 Changes audit — git status, uncommitted, TODOs
2 Quality check — lint, typecheck, tests
3 Learning capture — mistakes, patterns
4 Next session — blockers, what’s next
5 Summary — one paragraph handoff

12 Get Started

Install in Seconds

Cursor

/add-plugin pro-workflow

Claude Code

# Add marketplace
/plugin marketplace add rohitg00/pro-workflow

# Install plugin
/plugin install pro-workflow@pro-workflow

SkillKit (32+ Agents)

npx skillkit install pro-workflow
npx skillkit translate pro-workflow --agent cursor

Philosophy

Compound improvements — Small corrections → big gains
Trust but verify — Let AI work, review at checkpoints
Zero dead time — Parallel sessions keep momentum
Memory is precious — Both yours and Claude’s