pro-workflow

Pro Workflow

Complete AI coding workflow system from production use. Orchestration patterns, reference guides, and battle-tested habits that compound over time.

Works with: Claude Code, Cursor, Codex, Gemini CLI, and 32+ AI coding agents via SkillKit. Sections marked (Claude Code) use features specific to Claude Code — Cursor users can skip those or use the noted alternatives.

The Core Insight

"80% of my code is written by AI, 20% is spent reviewing and correcting it." — Karpathy

This skill optimizes for that ratio. Every pattern here reduces correction cycles.

---

1. The Self-Correction Loop

The single most powerful pattern. Your CLAUDE.md trains itself through corrections.

How It Works

When you correct Claude:

1. Claude acknowledges the mistake
2. Proposes a rule to prevent it
3. You approve → rule goes into memory
4. Future sessions avoid the same mistake

Add to CLAUDE.md

## Self-Correction Protocol

When the user corrects me or I make a mistake:
1. Acknowledge specifically what went wrong
2. Propose a concise rule: `[LEARN] Category: One-line rule`
3. Wait for approval before adding to LEARNED section

### LEARNED
<!-- Auto-populated through corrections -->

Trigger Phrases

• "Add that to your rules"
• "Remember this"
• "Don't do that again"

Example Flow

User: You edited the wrong file
Claude: I edited src/utils.ts when you meant src/lib/utils.ts.

[LEARN] Navigation: Confirm full path before editing files with common names.

Should I add this?

---

1b. Pre-Flight Discipline

Self-correction catches mistakes after the fact. This catches them before.

Karpathy's observations on LLM coding pitfalls name the upstream failures: silent assumptions, overcomplicated diffs, drive-by edits, vague success criteria. Four rules prevent each one.

Rule	Prevents
Surface, don't assume	Wrong interpretation, hidden confusion, missing tradeoffs
Minimum viable code	200-line diffs that should be 50, speculative abstractions
Stay in your lane	Drive-by refactors, "improvements" to adjacent code
Verifiable goals	Endless re-clarification, "make it work" loops

Full rules in rules/pre-flight-discipline.mdc (alwaysApply: true). Pairs with self-correction: pre-flight stops the mistake, self-correction captures the lesson when one slips through.

Add to CLAUDE.md

## Pre-Flight Discipline
Before coding: state assumptions, present ambiguity, push back if simpler exists.
Every changed line traces to the request - no drive-by edits.
Convert imperatives to verifiable goals: "fix bug" → "failing test → make it pass".

---

2. Parallel Sessions with Worktrees

Zero dead time. While one Claude thinks, work on something else.

Setup

Claude Code:

claude --worktree    # or claude -w (auto-creates isolated worktree)

Cursor / Any editor:

git worktree add ../project-feat feature-branch
git worktree add ../project-fix bugfix-branch

Background Agent Management (Claude Code)

• Ctrl+F — Kill all background agents (two-press confirmation)
• Ctrl+B — Send task to background
• Subagents support isolation: worktree in agent frontmatter

When to Parallelize

Scenario	Action
Waiting on tests	Start new feature in worktree
Long build	Debug issue in parallel
Exploring approaches	Try 2-3 simultaneously

Add to CLAUDE.md

## Parallel Work
When blocked on long operations, use `claude -w` for instant parallel sessions.
Subagents with `isolation: worktree` get their own safe working copy.

---

3. The Wrap-Up Ritual

End sessions with intention. Capture learnings, verify state.

/wrap-up Checklist

1. Changes Audit - List modified files, uncommitted changes
2. State Check - Run git status, tests, lint
3. Learning Capture - What mistakes? What worked?
4. Next Session - What's next? Any blockers?
5. Summary - One paragraph of what was accomplished

Create Command

~/.claude/commands/wrap-up.md:

Execute wrap-up checklist:
1. `git status` - uncommitted changes?
2. `npm test -- --changed` - tests passing?
3. What was learned this session?
4. Propose LEARNED additions
5. One-paragraph summary

---

4. Split Memory Architecture

For complex projects, modularize Claude memory.

Structure

.claude/
├── CLAUDE.md        # Entry point
├── AGENTS.md        # Workflow rules
├── SOUL.md          # Style preferences
└── LEARNED.md       # Auto-populated

AGENTS.md

# Workflow Rules

## Planning
Plan mode when: >3 files, architecture decisions, multiple approaches.

## Quality Gates
Before complete: lint, typecheck, test --related.

## Subagents
Use for: parallel exploration, background tasks.
Avoid for: tasks needing conversation context.

SOUL.md

# Style

- Concise over verbose
- Action over explanation
- Acknowledge mistakes directly
- No features beyond scope

---

5. The 80/20 Review Pattern

Batch reviews at checkpoints, not every change.

Review Points

1. After plan approval
2. After each milestone
3. Before destructive operations
4. At /wrap-up

Add to CLAUDE.md

## Review Checkpoints
Pause for review at: plan completion, >5 file edits, git operations, auth/security code.
Between: proceed with confidence.

---

6. Model Selection

Opus 4.6 and Sonnet 4.6 both support adaptive thinking and 1M-token context (as of 2025-08). The 1M context is available as a beta option (via the context-1m-2025-08-07 beta header); the default context window remains 200K. Sonnet 4.5 (200K context) has been retired from the Max plan in favor of Sonnet 4.6. See Models overview for current capabilities.

Task	Model
Quick fixes, exploration	Haiku 4.5
Features, balanced work	Sonnet 4.6
Refactors, architecture	Opus 4.6
Hard bugs, multi-system	Opus 4.6

Adaptive Thinking

Opus 4.6 and Sonnet 4.6 automatically calibrate reasoning depth per task — lightweight for simple operations, deep analysis for complex problems. No configuration needed. Extended thinking is built-in.

Add to CLAUDE.md

## Model Hints (as of 2025-08)
Opus 4.6 and Sonnet 4.6 auto-calibrate reasoning depth — no need to toggle thinking mode.
Use subagents with Haiku for fast read-only exploration, Sonnet 4.6 for balanced work.
Docs: https://docs.anthropic.com/en/docs/about-claude/models/overview

---

7. Context Discipline

200k tokens is precious. Manage it.

Rules

1. Read before edit
2. Compact at task boundaries
3. Disable unused MCPs (<10 enabled, <80 tools)
4. Summarize explorations
5. Use subagents to isolate high-volume output (tests, logs, docs)

Context Compaction

• Auto-compacts at ~95% capacity (keeps long-running agents alive)
• Configure earlier compaction: CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50
• Use PreCompact hooks to save state before compaction
• Subagents auto-compact independently from the main session

Good Compact Points

• After planning, before execution
• After completing a feature
• When context >70%
• Before switching task domains

---

8. Learning Log

Auto-document insights from sessions.

Add to CLAUDE.md

## Learning Log
After tasks, note learnings:
`[DATE] [TOPIC]: Key insight`

Append to .claude/learning-log.md

---

Learn Claude Code

Run /learn for a topic-by-topic guide covering sessions, context, CLAUDE.md, subagents, hooks, and more (see commands/learn.md). Official docs: https://code.claude.com/docs/

---

Quick Setup

Minimal

Add to your CLAUDE.md:

## Pro Workflow

### Self-Correction
When corrected, propose rule → add to LEARNED after approval.

### Planning
Multi-file: plan first, wait for "proceed".

### Quality
After edits: lint, typecheck, test.

### LEARNED

Full Setup

git clone https://github.com/rohitg00/pro-workflow.git /tmp/pw
cp -r /tmp/pw/templates/split-claude-md/* ./.claude/
cp -r /tmp/pw/commands/* ~/.claude/commands/

---

Hooks (Claude Code)

Pro-workflow includes automated hooks to enforce the patterns. Cursor users get equivalent enforcement through .mdc rules in the rules/ directory.

PreToolUse Hooks

Trigger	Action
Edit/Write	Track edit count, remind at 5/10 edits
git commit	Remind to run quality gates
git push	Remind about /wrap-up

PostToolUse Hooks

Trigger	Action
Code edit (.ts/.js/.py/.go)	Check for console.log, TODOs, secrets
Test commands	Suggest [LEARN] from failures

Session Hooks

Hook	Action
SessionStart	Load LEARNED patterns, show worktree count
Stop	Context-aware reminders using last_assistant_message
SessionEnd	Check uncommitted changes, prompt for learnings
ConfigChange	Detect when quality gates or hooks are modified mid-session

Install Hooks

# Copy hooks to your settings
cp ~/skills/pro-workflow/hooks/hooks.json ~/.claude/settings.local.json

# Or merge with existing settings

Hook Philosophy

Based on Twitter thread insights:

• Non-blocking - Hooks remind, don't block (except dangerous ops)
• Checkpoint-based - Quality gates at intervals, not every edit
• Learning-focused - Always prompt for pattern capture

---

Contexts

Switch modes based on what you're doing.

Context	Trigger	Behavior
dev	"Let's build"	Code first, iterate fast
review	"Review this"	Read-only, security focus
research	"Help me understand"	Explore, summarize, plan

Use: "Switch to dev mode" or load context file.

---

Agents

Specialized subagents for focused tasks.

Agent	Purpose	Tools
planner	Break down complex tasks	Read-only
reviewer	Code review, security audit	Read + test

When to Delegate

Use planner agent when:

• Task touches >5 files
• Architecture decision needed
• Requirements unclear

Use reviewer agent when:

• Before committing
• PR reviews
• Security concerns

Custom Subagents (Claude Code)

Create project-specific subagents in .claude/agents/ or user-wide in ~/.claude/agents/:

• Define with YAML frontmatter + markdown system prompt
• Control tools, model, permission mode, hooks, and persistent memory
• Use /agents to create, edit, and manage interactively
• Preload skills into subagents for domain knowledge

Agent Teams (Claude Code, Experimental)

Coordinate multiple Claude Code sessions as a team:

• Enable: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
• Lead session coordinates, teammates work independently
• Teammates message each other directly (not just report back)
• Shared task list with dependency management
• Display: in-process (Shift+Down to navigate, wraps around) or split panes (tmux/iTerm2)
• Delegate mode (Shift+Tab): lead coordinates only, no code edits
• Best for: parallel reviews, competing hypotheses, cross-layer changes
• Docs: https://code.claude.com/docs/agent-teams

---

9. Orchestration: Command > Agent > Skill

The most powerful pattern for complex features. Three layers, each with a single job.

The Architecture

Command (user-facing entry point)
  └── Agent (execution, constrained tools, preloaded skills)
        └── Skill (domain knowledge, injected at startup)

Multi-Phase Development (/develop)

For features touching >5 files or needing architecture decisions:

1. Research → orchestrator agent explores codebase, scores confidence (0-100)
2. Plan → presents approach, files to change, risks. Waits for approval.
3. Implement → executes plan step by step with quality gates every 5 edits
4. Review → reviewer agent checks for security, logic, quality

All four phases run in order. Each phase requires explicit user approval before the next phase begins.

Agent Skills (Preloaded)

# Agent frontmatter
skills: ["api-conventions", "project-patterns"]

Full skill content injected at agent startup. Use for knowledge the agent always needs.

On-Demand Skills (Invoked)

Skills with user-invocable: true are called via /skill-name. Use context: fork for isolated execution that doesn't pollute main context.

When to Orchestrate

Scenario	Pattern
Feature > 5 files	/develop with orchestrator
Bug investigation	debugger agent
Quick exploration	scout agent (background)
Code review	reviewer agent
Simple task	Just do it directly

---

10. Daily Habits

Every Session

• Run /doctor if things feel off
• Manual /compact at 50% — don't wait for auto-compact
• ultrathink in prompts for maximum reasoning
• Name sessions with /rename for easy /resume
• End with /wrap-up to capture learnings

Context Management

• CLAUDE.md: < 60 lines root, < 150 max
• Use CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50 for proactive compaction
• Delegate heavy exploration to subagents
• Keep <10 MCPs, <80 tools

Cross-Agent Tips

• Use Cursor for tab completions + Claude Code in terminal for hard problems
• Same MCP servers work across both (share .mcp.json at project root)
• SkillKit translates skills to any agent: npx skillkit translate pro-workflow --agent cursor

---

MCP Config (Claude Code)

Start with 3 MCPs. Add only for concrete needs.

Essential:

• context7 — Live documentation lookup
• playwright — Browser automation (most token-efficient)
• github — PRs, issues, code search

See mcp-config.example.json for setup and curated recommendations.

---

Commands (Claude Code)

Command	Purpose	Cursor Equivalent
/wrap-up	End-of-session ritual	wrap-up skill
/learn-rule	Extract correction to memory	learn-rule skill
/develop	Multi-phase feature build	orchestrate skill
/doctor	Health check	—
/commit	Smart commit with quality gates	smart-commit skill
/insights	Session analytics and patterns	insights skill
/replay	Surface past learnings	replay-learnings skill
/handoff	Session handoff document	session-handoff skill
/search	Search learnings by keyword	—
/list	List all stored learnings	—
/learn	Topic-by-topic Claude Code guide	—

---

Reference Guides

Deep dives on configuration and features:

Guide	Topics
docs/settings-guide.md	All settings keys, permission modes, hierarchy, sandbox, env vars
docs/cli-cheatsheet.md	Every CLI flag, keyboard shortcut, slash command
docs/orchestration-patterns.md	Command > Agent > Skill architecture, frontmatter reference
docs/context-loading.md	CLAUDE.md monorepo loading, agent memory, skills discovery
docs/cross-agent-workflows.md	Claude Code + Cursor config mapping, background agents
docs/new-features.md	Voice mode, agent teams, checkpointing, new hook events
docs/daily-habits.md	Session habits, debugging tips, terminal setup, anti-patterns

---

Philosophy

1. Compound improvements - Small corrections lead to big gains
2. Trust but verify - Let AI work, review at checkpoints
3. Zero dead time - Parallel sessions keep momentum
4. Memory is precious - Yours and the AI's
5. Orchestrate, don't micromanage - Wire patterns together, let agents execute

---

Complete AI coding workflow system from production use across Claude Code, Cursor, and beyond.