Claude Code
Mini Tutorial
How to use Claude Code to be 100x more productive. Complete system guide for knowledge workers.
Who's Teaching You This
Iwo Szapar
Co-founded AI Maturity Index (acquired by ISG, Nasdaq: III in January 2026). Helped 3,000+ companies transform how they work. Featured in Financial Times and Forbes—and I run my entire business on the system you're about to learn.
Credentials:
- • 1,164+ days of daily AI usage
- • 30+ custom agents in production
- • Built 3 AI-powered products
- • Consultant to Fortune 500 teams
What Makes Me Different:
- • Context engineering, not prompt tricks
- • Real system with 30+ agents (not theory)
- • Non-technical focus
How to Navigate This Tutorial
This guide is structured in 6 progressive stages, plus a new harness engineering layer for reliable AI work.
Stages 1-2: Foundations (15 min)
Understand what Claude Code is and get it installed
Stages 3-4: Core Power (30 min)
Progressive disclosure, hooks, agents, skills, and compound learning
Stage 5: Advanced (20 min)
MCP integrations for connecting your business tools
Stage 6: Implementation (10 min)
Your path forward with pre-built templates
Pro Tip: Click section titles to expand/collapse. Save time by focusing on what you need most.
Quick Start: Your First 10 Minutes
Want to see the magic before reading the theory? Pick your setup and do this right now.
Install Claude Code (2 min)
curl -fsSL https://claude.ai/install.sh | bashOpen terminal, authenticate, start
claude login
claudeSay this (literally copy-paste):
"Research the top 3 competitors in my space and
summarize their pricing models. Run them in parallel
with subagents."Watch Claude spawn multiple agents automatically
You just ran a multi-agent workflow. Three AI workers researched in parallel while you sat back. That's the core idea—now read on for how to build a full system around it.
Already using Claude Code? Skip to Stage 4: Agents, Skills, Hooks & Harnesses for the advanced orchestration patterns.
Understanding: The Shift from Chat to Repository-Based AI
30-Second Intro
Claude Code is an AI coding agent by Anthropic that works across your terminal, IDE, desktop app, and browser. Sessions still start fresh, but your repository can carry durable context through files like CLAUDE.md, .claude/rules/, skills, agents, and project docs.
"Claude Code is the most underrated AI tool for non-technical people."
Three Ways to Work with Claude
The naming is confusing because the desktop app now contains multiple tabs. Separate the product surface from the workflow you want.
Chat tab · Browser or app
Claude Chat
- • Best for questions, drafting, brainstorming
- • No repository workflow required
- • Files are attachments, not a working project brain
- • Use this when the task is one-off
Best for: fast thinking, not reusable systems
Cowork tab · Desktop app
Claude Cowork
- • Autonomous background agent in a cloud VM
- • Good for longer knowledge-work tasks
- • Lower setup than terminal workflows
- • Great starting point if you want outcomes, not infrastructure
Best for: consultants, analysts, operators
Code tab, CLI, IDE, web
Claude Code (this guide)
- • Coding agent for files, repos, GitHub, and automation
- • Runs in terminal, IDEs, desktop app, browser, Slack, and CI/CD
- • Skills, subagents, hooks, MCP, routines, and background sessions
- • Most powerful when your work lives in a structured repository
Best for: power users building a reusable operating system
You're reading the guide for this ↗Not sure? Use Chat for one-off thinking, Cowork for low-setup background work, and Claude Code when you want project memory, custom workflows, hooks, MCP, and repeatable systems.
Any knowledge worker who:
- Does repeated work that follows patterns (reports, emails, proposals, content)
- Manages ongoing projects where context matters across weeks and months
- Juggles multiple tools (CRM, email, calendar, analytics, content platforms)
- Wants AI that gets smarter over time, not one that forgets everything after each chat
Consultants
Client reports, proposals, frameworks
Product Managers
PRDs, roadmaps, user stories
Marketers
Content calendars, campaigns
Founders
Strategy, investor updates, ops
Sales & Revenue
Pipeline, outreach, CRM automation
Researchers & Analysts
Data synthesis, competitive intel
Not ideal for:
One-off questions with no follow-up, simple brainstorming sessions, or tasks that never repeat. For those, regular Claude Chat works fine.
"I'm not a coder. I don't write code. So it's incredible that this type of stuff is available and can happen."
— Dominic, Cloud Consultant • Built a subscription platform using Claude Code with zero coding background
Before we dive in
See What You're Building Toward
This tutorial teaches each concept step by step. But if you want to see what a complete system looks like in practice, here's a 3-minute clip from a live session.
One command. AI reads your emails, WhatsApp messages, tasks, and CRM — then tells you what to focus on today. Everything you see here, you'll learn how to build in the stages below.
Problem #1: Zero Persistent Memory
Traditional AI (ChatGPT, Claude Chat): Every conversation starts from scratch. Re-explain your project 10+ times per week.
Claude Code Solution: Repository as Persistent Memory. CLAUDE.md file loads at every session start. Dev Docs (plan.md, context.md, tasks.md) for multi-week projects.
Traditional AI Session:
Session 1:
You: "I'm building a SaaS product..."
[Explain project for 10 minutes]
Session 2 (next day):
You: "Remember my SaaS product?"
AI: "What product? Please explain."
[Re-explain AGAIN]Claude Code Session:
Session 1:
Claude reads CLAUDE.md
"I see you're building a SaaS
product. What would you like?"
Session 2 (next day):
Claude reads CLAUDE.md
"Back to the SaaS project.
What's next?"Teaching Point: Persistent memory eliminates context-switching overhead
Real Impact: Richard (management consultant, Australia) spent a week setting up Claude Code independently. After one 2-hour session with Second Brain templates: "You've solved for me on this call what I was still doing wrong—I was prompting it like it was a chatbot. Whereas by using commands and agents, you can effectively have that manage your daily workflow for you. That's a huge thing. Definitely worth the money to save me three to four months just hacking my way through it."
Problem #2: Sequential Bottleneck
Traditional AI: Can only work on ONE thing at a time. Analyze 10 files? Wait for each sequentially.
Claude Code Solution: Subagents run multiple tasks in parallel inside the session. For independent long-running work, use background sessions from Claude Code web, desktop, Slack, or Agent View.
SEQUENTIAL (ChatGPT Way):
[Agent 1: Research] → [Agent 2: Content] → [Agent 3: Analysis]
15 min 20 min 10 min
Total: 45 minutes
PARALLEL (Claude Code Way):
[Agent 1: Research] ─┐
[Agent 2: Content] ─┼─ All running simultaneously
[Agent 3: Analysis] ─┘
Total: MAX(15, 20, 10) = 20 minutes
Speedup: 45 min → 20 min = 2.25x fasterReal Impact: A consultant analyzed 5 competitors (Asana, Monday, ClickUp, Notion, Airtable). ChatGPT took 45 min and only completed 3. Claude Code with subagents took 20 min and completed all 5 simultaneously.
Teaching Point: Background agents enable true parallel processing
Real Impact: Damian (AI consultant, Switzerland) built a complete social media content flow in 2 days—something a commercial SaaS tool sells for hundreds/month. His workflow: capture ideas → extract LinkedIn voice DNA → generate carousel PDFs → schedule posts. "A friend just told me about this new super cool solution he bought. It was what I built in two days with the Second Brain. Pure magic."
Problem #3: Context Switching Tax
Traditional AI: Copy docs → Paste → Wait. Copy code → Paste → Wait. 30+ minutes lost per session.
Claude Code Solution: MCP (Model Context Protocol) connects directly to Gmail, Calendar, CRM, Analytics. No more copy-paste between apps.
Manual Process:
- 1. Open HubSpot
- 2. Export 200 webinar attendees
- 3. Download CSV
- 4. Copy email list
- 5. Paste into ChatGPT
- 6. Ask for analysis
- 7. Copy results back
- 8. Upload to HubSpot
Time: 35 minutes
MCP Process:
- 1. "Tag webinar attendees"
- 2. Claude connects to HubSpot
- 3. Done.
Time: 3 minutes
Real Impact: Sales Ops manager went from 2 hours/week on manual CRM updates to zero. All automated via MCP.
Problem #4: Tool Isolation
Traditional AI: AI can't access your actual business tools. Manual data transfer everywhere.
Claude Code Solution: Skills auto-activate based on keywords. Hooks run scripts automatically. Build once, reuse forever.
Traditional AI:
You: "Generate weekly sales report"
AI: "Please provide data..."
You: [Copy sales from HubSpot]
You: [Copy traffic from Analytics]
You: [Copy revenue from Stripe]
AI: [Generates report]
You: [Format manually]
You: [Send to team]
Time: 45 minutes
Claude Code with Skills:
You: "Generate weekly sales report"
Skill auto-loads:
├─ MCP: Fetch HubSpot deals
├─ MCP: Fetch GA4 traffic
├─ MCP: Fetch Stripe revenue
└─ Generate formatted report
└─ Save to reports/weekly-2025-01-11.md
Time: 2 minutesReal Impact: Marketing manager automated 5 weekly reports. Went from 3.5 hours/week to 10 minutes/week. That's 21x faster.
What One Tool + Structured Context Actually Replaced
This isn't theoretical. These are real admin pages running in production right now.
| SaaS Category | Typical Monthly Cost | Now Built Into Second Brain |
|---|---|---|
| CRM (HubSpot/Pipedrive) | $45-100/mo | Sales pipeline + prospect tracking |
| Email Marketing (Mailchimp) | $30-60/mo | Automated sequences + newsletter |
| Content Scheduling (Buffer) | $15-30/mo | Content calendar + queue management |
| Task Management (Asana/Linear) | $10-25/mo | Task tracking + session workflow |
| Analytics Dashboard | $20-50/mo | Business intelligence + reporting |
| Client Delivery Portal | Custom | Repository generator + onboarding |
$120-265/mo before
$20/mo after (Claude Pro)
This isn't a comparison between Claude and HubSpot. HubSpot is better at being a CRM. But for a solopreneur or small team, Claude Code + structured context handles 80% of what you actually use those tools for — at 10% of the cost.
Stage 1 Checklist
Want the Shortcut?
Skip 2-3 weeks of trial & error. Everything below is pre-built in the Second Brain.
See PackagesSetup: Getting Started
Claude Subscription or API Billing
Most people should start with a Claude Pro or Max plan for predictable monthly usage. API billing is also available for programmatic or team workflows.
Choose How to Use Claude Code
Choose the surface that matches your comfort level and operating system
Claude Code on the Web
Research preview. Connect GitHub at claude.ai/code. Runs in an Anthropic-managed cloud VM.
Desktop App Code Tab
macOS and Windows native app. GUI for Claude Code with sidebar, terminal, file editor, diffs, previews, and PR monitoring.
Terminal CLI (Most Powerful)
Fullest control over local files, shell, MCP, hooks, scripts, CI, and automation.
IDE Extension
VS Code, Cursor, and JetBrains-family IDEs. Claude Code runs inside your editor alongside your files.
GitHub Account (Optional but Recommended)
Think of it as "Google Drive for your AI's brain" - backs up your repository, syncs across devices.
Stage 2 Checklist
The Best Way to Adopt Claude Code
Don't try to build the whole system at once. The biggest mistake people make is trying to automate everything on day one. That never works.
Pick one use case per day. That's it.
Day 1
"Help me research this company."
Day 2
"Write me a LinkedIn post about X."
Day 3
"Audit this landing page."
Day 4
"Create a skill for my weekly report."
Little by little, your repo fills up. Your skills, agents, rules, and CLAUDE.md get smarter. After a few weeks you look back and realize you've built an entire system—one small piece at a time.
Your 4-Week Adoption Roadmap
Sub-Agents (Zero Setup Needed)
Ask Claude Code to use subagents for 2-3 independent research tasks. Notice how your main conversation stays clean while specialists work in parallel.
Your First Skill
Pick your most repeated workflow (content creation, code review, data analysis). Create .claude/skills/my-workflow/SKILL.md with a clear description and step-by-step instructions. Test by asking naturally or invoking /my-workflow, then iterate.
Skills + Sub-Agents Combined
Inside your skill, add steps that use sub-agents for research. Chain two skills together. Set up your CLAUDE.md with project context and preferences.
Harness + MCP + Advanced (Optional)
Add plan, test, review, and approval gates around your most important workflows. Then connect business tools via MCP. Most people should build the harness before adding more autonomy.
Memory System: CLAUDE.md & Repository Structure
Note: CLAUDE.md is optional but strongly recommended for persistent context. It eliminates re-explaining your project every session.
A special markdown file that Claude Code loads at the start of every session. Think of it as your project's instruction manual - but for AI, not humans.
Without CLAUDE.md:
- • Every session: "Here's my project..."
- • Re-explain context 10+ times/week
- • Claude forgets your preferences
- • Inconsistent outputs
With CLAUDE.md:
- • Claude reads once per session
- • Instant context about your project
- • Remembers style & requirements
- • Consistent quality
# CLAUDE.md
## Project Overview
[One sentence: What you're building]
**Mission**: [Your project's goal]
**Target Users**: [Who uses this]
## Repository Structure
- `company-brain/`: Business knowledge
- `agents/`: AI agent templates
- `data/`: Datasets and analysis
## Key Guidelines
- [Important fact #1 Claude should always know]
- [Important fact #2]
- [Important fact #3]
## Output Standards
- [How you want deliverables formatted]
- [Your writing style preferences]
## Guardrails
- [Things Claude should NEVER do]
- [Data privacy rules]See it in practice
What a Real Second Brain Looks Like on GitHub
Forget the theory for a moment. Here's an actual Second Brain repository with 30+ agents, skills, commands, and knowledge files — the structure that makes everything in this tutorial work.
"I was putting everything in CLAUDE.md—500 lines. Claude started ignoring half of it. Then I learned about progressive disclosure and everything clicked."
— Andrew, Product Director
The #1 Mistake Everyone Makes
People cram everything into CLAUDE.md—business context, API docs, style guides, gotchas, templates. At 500+ lines, Claude starts ignoring instructions. Research shows instruction-following degrades past ~300 lines and ~50 rules. More instructions paradoxically leads to worse instruction-following.
The solution is progressive disclosure: keep CLAUDE.md short (under 200 lines) with pointer tables that load detailed docs on demand. Think of it as a table of contents, not an encyclopedia.
Monolithic (Breaks at Scale):
CLAUDE.md (972 lines)
├─ Business context (50 lines)
├─ API documentation (200 lines)
├─ Database schema (150 lines)
├─ Style guide (100 lines)
├─ Email templates (120 lines)
├─ Gotchas list (80 lines)
├─ MCP configs (150 lines)
└─ ... Claude ignores half of itLayered (Scales Infinitely):
CLAUDE.md (200 lines)
├─ Business context (always loaded)
├─ Tech stack (always loaded)
├─ Pointer table → docs/
├─ Critical gotchas (3-5 rules)
└─ Self-improvement protocol
.claude/docs/ (loaded on demand)
├─ api-guide.md
├─ database-ops.md
├─ email-system.md
└─ gotchas.mdThe Three Layers
Layer 1: CLAUDE.md (Always Loaded — Under 200 Lines)
Only universally applicable context belongs here. Business context, tech stack, critical gotchas, and a pointer table to everything else.
## Architecture & Patterns
| Area | Read This File |
|-------------------|----------------------------|
| Writing API code | .claude/docs/api-guide.md |
| Database work | .claude/docs/database.md |
| Email system | docs/EMAIL_SYSTEM.md |
| Known gotchas | .claude/docs/gotchas.md |
| Troubleshooting | docs/TROUBLESHOOTING.md |
Read the relevant file BEFORE making changes
in that area.When Claude needs to work on the email system, it reads CLAUDE.md, sees the pointer, and loads docs/EMAIL_SYSTEM.md on demand. The other docs stay out of context.
Layer 2: .claude/docs/ (Loaded On Demand)
Detailed documentation that Claude reads only when working in a specific area. Each file starts with a "When to read" trigger.
# API Development Guide
> When to read: Before writing or modifying
> any API endpoint
## Pattern: Legacy Request/Response
All endpoints MUST use VercelRequest/
VercelResponse pattern. Web API pattern
causes 30-second timeouts.
## Required: .js Extensions
All imports in API files MUST include .js
extension. Missing extensions cause silent
build failures in production.
## Template
export default async function handler(
req: VercelRequest,
res: VercelResponse
) { ... }Layer 3: knowledge/ (Agent-Specific Context)
Deep reference material that only specific agents or skills need. Never loaded into general sessions.
.claude/knowledge/
├─ content-creator/ # Voice DNA, templates
│ ├─ brand-voice.md
│ └─ content-types.md
├─ sales/ # CRM context, scripts
│ ├─ objection-handling.md
│ └─ email-sequences.md
└─ data-analysis/ # Schema, query patterns
└─ common-queries.mdWhy This Works:
Context Efficiency
A lean CLAUDE.md with pointers uses far fewer tokens than one giant file. Leaves room for actual work.
Better Instruction Following
50 rules in a focused file beat 200 rules in a bloated file. Every time.
Scales Infinitely
Add new docs without touching CLAUDE.md. Just add a row to the pointer table.
Quick Start: Progressive Disclosure in 3 Steps
- 1. Create .claude/docs/ directory
Move any section longer than 20 lines out of CLAUDE.md into its own file.
- 2. Add pointer table to CLAUDE.md
Map each area of work to the file Claude should read before starting.
- 3. Add "When to read" triggers
Start each docs file with a one-line trigger: "When to read: Before working on [X]"
Building This Architecture Takes 4-6 Hours
Writing pointer tables, organizing docs with "when to read" triggers, testing that Claude loads the right file at the right time. The Second Brain ships with 10+ pre-configured docs, trigger tables, and the full progressive disclosure architecture ready to go.
Skip the SetupStage 3 Checklist
Automation: Agents, Skills, Hooks, Harnesses & Learning
"That's the real advantage of Claude Code. Parallel programming. You fire off 3 research agents and they all come back with results while you keep working on the main thread."
— Wytze, Community Member
Subagents are specialized AI workers with isolated context. Like having a research analyst AND a writer on your team working simultaneously.
When to Use Subagents:
- ✅ Parallel analysis (10 competitors at once)
- ✅ Specialized expertise needed
- ✅ Independent subtasks
- ✅ Large projects with distinct phases
When NOT to Use:
- ❌ Single simple tasks
- ❌ Sequential dependencies
- ❌ Tasks requiring shared context
Real Example: Competitive Analysis
Task: Analyze 5 competitors (Asana, Monday, ClickUp, Notion, Airtable)
Sequential (ChatGPT): 45 minutes
Analyze one → Next → Next... 3 done poorly, 2 skipped due to fatigue
Parallel (Claude Code): 8 minutes
5 subagents run simultaneously. All 5 complete, high quality, consistent format
USER PROMPT:
"Analyze these 5 competitors: Asana, Monday, ClickUp,
Notion, Airtable. For each, analyze pricing, features,
target market."
CLAUDE CODE SPAWNS 5 SUBAGENTS:
┌─────────────────────────────────────────────┐
│ Main Claude │
│ ├─ Subagent 1: Analyze Asana │
│ ├─ Subagent 2: Analyze Monday │
│ ├─ Subagent 3: Analyze ClickUp │
│ ├─ Subagent 4: Analyze Notion │
│ └─ Subagent 5: Analyze Airtable │
│ │
│ [All run in parallel] │
│ │
│ After 8 min: Aggregate results → │
│ Generate comparison table │
└─────────────────────────────────────────────┘Cost Reality Check
Sub-agents consume their own context windows. A 3-agent parallel task uses roughly 3x the tokens of doing it sequentially. Agent teams (5+ agents) can hit 7x or more. This matters if you're on API billing.
| Strategy | How | Savings |
|---|---|---|
| Smaller models for workers | Use Haiku for simple tasks, Opus for decisions | 60-80% |
| Limit agent scope | Restrict tools and file access per agent | 30-50% |
| Use Explore agents for search | Read-only, fast, minimal token usage | ~70% |
| Cache in CLAUDE.md | Store results so agents don't re-research | Significant |
| Set max_turns | Prevent runaway agent loops | Hard cap |
On a Claude Pro/Max subscription ($20-200/mo), sub-agents run within your plan limits—no extra per-token cost. API key users should monitor usage more carefully.
See it in practice
What's Actually Inside an Agent File
You just read about agent configs. Here's one being opened live — a content creator agent with tool access, a decision tree, voice settings, and references to knowledge files. It's a text file, not code.
"I've noticed better content when I have a central agent then a suite of skills it can invoke on its own. Instead of me remembering the exact workflow every time, I just say 'write a LinkedIn post' and it pulls the right skill."
— Shaun, Marketing & Growth Specialist
Agent = Model + Harness
Most people keep upgrading the model. Power users upgrade the harness around the model. A harness is the operating system for AI work: it defines the plan, gives the agent safe tools, checks the output, asks for review, captures lessons, and only then lets the work ship.
The model is only one input. The harness is everything else: prompts, CLAUDE.md, rules, tools, MCP servers, filesystem, Git, sandboxes, hooks, subagents, context policies, logs, graders, recovery paths, and release gates. Claude Code, Cursor, Codex, Aider, Cline, and Managed Agents are all harnesses around models.
The Ratchet: Every Failure Becomes a Permanent Fix
Harness engineering starts when you stop blaming the model by default. If the agent repeats a mistake, the job is to change the harness so that exact failure is harder or impossible next time.
| Observed failure | Harness fix | Where it lives |
|---|---|---|
| Ignored a project convention | Add a short earned rule | CLAUDE.md or .claude/rules/ |
| Ran a destructive command | Block it before execution | PreToolUse hook |
| Got lost in a long task | Split planner, executor, and reviewer roles | Skill + subagents |
| Finished with broken output | Run objective checks and feed failures back | Verification hook or grader |
| Repeated the same weak draft | Add examples, rubric, and review questions | Skill reference files |
Add constraints only when they trace back to a real failure. Remove them when they are obsolete. A good harness is a living artifact, not a prompt graveyard.
1. Plan Gate
Before Claude acts, it must restate the goal, assumptions, files or sources it will touch, and what "done" means.
2. Work Boundary
The agent gets only the tools and scope it needs. Read-only for research. Approval before sending, deleting, publishing, or changing records.
3. Verification Gate
The output is checked against objective criteria: sources cited, numbers reconciled, checklist complete, required fields present.
4. Review Gate
A separate pass looks for the things automation misses: wrong diagnosis, overengineering, tone mismatch, hidden assumptions.
5. Compound Gate
Every mistake becomes a rule, template, example, or checklist item. The system gets better because the harness learns.
6. Ship Gate
Nothing leaves the system until the plan, verification, review, and approval artifacts exist. Claude can propose; the harness decides.
The Load-Bearing Parts of a Harness
State
Filesystem, Git, task files, memory files, and output folders. This is where work survives beyond the current context window.
Tools
Bash, code execution, browser automation, MCP servers, and APIs. Keep tools focused; overlapping tools confuse the model and inflate context.
Context Policy
What gets loaded, compacted, offloaded, or retrieved later. This decides what crosses the model's context boundary.
Enforcement
Hooks, permissions, sandboxes, test commands, and approval gates. These turn preferences into constraints.
Evaluation
Rubrics, graders, review agents, and acceptance criteria. Grade the outcome, not the weird path the agent took.
Observability
Logs, traces, token usage, cost, latency, and failure clusters. A score with no ticket is just dashboard art.
Context Management Is Harness Design
The model cannot think about tokens it never receives. When the context window fills up, the harness must decide what stays, what gets compressed, what moves to disk, and what is retrieved later. That decision is often more important than the model choice.
Truncate
Drop low-value output or old transcript when continuity does not matter.
Compact
Summarize the session and continue with the compressed state.
Offload
Save long logs, tool results, and drafts to files; keep only the useful preview in context.
Disclose
Load detailed docs, skills, and tools only when the task actually needs them.
Practical rule: keep always-loaded context short, page through large files with offset/limit, use search before reading whole files, and store long tool outputs on disk instead of pasting them back into the conversation.
Outcome Rubrics: Define What Done Looks Like
Anthropic's Managed Agents API makes this pattern explicit: define an outcome, attach a rubric, let a separate grader evaluate the artifact, then feed gaps back to the agent for another iteration. You can use the same idea manually in Claude Code.
Outcome:
Create a competitor pricing report for the 5 named companies.
Rubric:
- Includes all 5 companies
- Every price claim has a source URL
- Pricing tiers are normalized into one table
- Unknown prices are marked "not published" instead of invented
- Final section recommends what we should change
Iteration rule:
If any rubric item fails, revise once before showing me the final.The important move is separation: one loop creates the artifact, another evaluates it against the rubric. This reduces self-congratulation and catches gaps faster.
The Knowledge Worker Harness Template
1. Track
Capture the request in a task, client folder, or project log.
2. Plan
Ask Claude: "Before doing the work, give me the plan,
assumptions, sources/tools needed, and acceptance criteria."
3. Work
Let Claude execute inside clear boundaries:
read-only research, draft-only writing, no external sends.
4. Verify
Run a checklist that matches the work type:
sources, calculations, brand voice, missing fields, risks.
5. Review
Ask a second pass: "Find what is wrong, overbuilt,
unsupported, or not how we do things here."
6. Learn
Capture one reusable rule or template improvement.
7. Ship
Only publish, send, update CRM, or hand off after approval.Copy-Paste Prompt: Turn Any Workflow Into a Harness
I want to turn this repeated workflow into a reliable AI harness:
[describe the workflow]
Create:
1. The planning questions Claude must answer before starting
2. The tool and permission boundaries
3. The verification checklist
4. The review questions a second pass should ask
5. The human approval points
6. The learning rule to update after each run
Keep it practical for a non-technical knowledge worker.Weekly Harness Audit: Cut the Invisible Overhead
If Claude feels slower, dumber, or burns through limits, audit the harness before blaming the model. Most waste hides in always-loaded context.
| Check | What to look for | Fix |
|---|---|---|
CLAUDE.md size | Rules that no longer matter or belong in narrower docs | Keep root context short; move specifics to rules, docs, or skills |
| Hooks | UserPromptSubmit or SessionStart hooks that inject context every time | Disable anything without a specific job |
| MCP servers | Tool schemas loaded for tools you rarely use | Keep only daily tools always-on; enable others per session |
| Skills | Broad descriptions causing irrelevant skills to load | Tighten descriptions or disable unused skills |
| Conversation length | Long sessions re-reading stale history | Compact, summarize, or start fresh with a handoff note |
Examples by Role
| Workflow | Verification Gate | Human Gate |
|---|---|---|
| Consulting report | Claims tied to client data, recommendations mapped to scope, no confidential names leaked | Partner reviews insight quality before sending |
| Sales outreach | CRM history checked, personalization source cited, no invented trigger event | You approve before email, LinkedIn, or WhatsApp send |
| Financial analysis | Totals reconcile, assumptions listed, anomalies flagged, source files named | You approve interpretation before sharing |
| Content publishing | Voice check, banned phrases removed, facts sourced, CTA present | You approve final version before queue or publish |
What No Harness Reliably Catches
- • Misdiagnosis: solving the wrong root cause
- • Overengineering: adding complexity nobody asked for
- • Misunderstood instructions: doing the adjacent task
- • "We don't do it that way here" team norms
- • Taste, judgment, and executive tradeoffs
- • Social accountability for the final outcome
The rule: automate the typing, research, checking, and formatting. Keep responsibility, final judgment, and external commitments with a human.
"I'm in LOVE with this new system. I did about 10 days of work in a single day."
— Shaun, Marketing & Growth Specialist • Built audio-to-social-post automation in his first week
This Is What 5,475 People Commented "BRAIN" For
Most people use Claude as a single assistant. Ask a question, get an answer. But Claude Code lets you build an entire team of specialists that hand off work between themselves—while you review.
Important distinction: subagents vs background agents
Subagents are specialists inside one Claude Code session. They have their own context and tool permissions, then report back to the main session. Background agents are separate cloud sessions you manage through Claude Code on web, desktop, Slack, or the Agent View.
Use subagents when:
One main workflow needs several focused analyses, reviews, or drafts in parallel.
Use background agents when:
You want independent long-running sessions on different repos, branches, PRs, or scheduled jobs.
Three Levels of Agent Coordination
Level 1: Solo Agent
One Claude session doing everything. Like working with one very capable generalist. This is where most people stop.
Level 2: Delegating Lead (Subagents)
Your main Claude session acts as lead. It dispatches specialists to do focused work—research, drafting, analysis—in parallel. Each specialist reports back to the lead, which synthesizes results. This is production-ready and what most power users build.
Level 3: Background Sessions + Agent View
Multiple Claude Code sessions running independently. Use claude agents or the web/desktop surfaces to dispatch, monitor, and resume work. Research-preview territory—powerful but still evolving.
Real Example: Sales Outreach Pipeline
Here's what actually happens when I say "draft outreach for this prospect":
Lead Agent orchestrates the pipeline:
│
├─ Research Agent
│ → Pulls LinkedIn profile + company intel
│ → Returns: role, company size, recent posts
│
├─ CRM Agent
│ → Checks prospect history, stage, past notes
│ → Returns: last contact, deal stage, context
│
├─ Draft Agent
│ → Writes personalized email using research
│ → Uses voice guidelines from knowledge files
│
├─ Quality Agent
│ → Scores draft against brand voice rubric
│ → Flags generic phrases, suggests improvements
│
└─ Lead Agent
→ Reviews final draft
→ I approve → sends via email MCP
5 agents. 3 minutes.
What used to take 45 minutes of tab-switching
between LinkedIn, CRM, email, and style guides.Multi-Agent Review: When Specialists Challenge Each Other
With Level 2, all communication flows through the lead session. You can still ask Claude to run independent specialist passes, compare their findings, and synthesize the disagreements. For separate long-running workstreams, use background sessions and Agent View.
"Our conversion rate dropped 15% last week.
Use three specialist subagents to investigate:
- One analyzes traffic sources
- One reviews checkout funnel
- One checks pricing page changes
Then compare their findings and challenge contradictions."
What happens:
Traffic Agent: "Paid traffic quality dropped"
Checkout Agent: "But checkout completion is
stable. The issue is upstream."
Pricing Agent: "New pricing page went live
Tuesday. Bounce rate up 40%."
Traffic Agent: "That explains the drop—
the traffic was fine, the page wasn't."
→ Root cause found in minutes, not hours.The debate structure prevents anchoring bias. Sequential investigation tends to find one theory and stop. Multiple independent investigators challenging each other surface the real answer.
Building 30+ specialized agents with the right personas, tool permissions, knowledge files, and coordination logic took months of iteration. Each agent needs to know what it can access, what format to output, and when to hand off to the next agent. The Second Brain ships with the full agent library—pre-configured for content creation, sales outreach, client management, data analysis, and daily operations.
Claude Code Works Alongside Other Tools
The most effective setups aren't "Claude Code only." Power users combine tools strategically: ChatGPT for creative brainstorming, Claude Code for structured execution, Gemini for review. Claude Code becomes the orchestration layer that ties everything together—not the only tool in your stack.
1. Brainstorm with ChatGPT (creative, divergent thinking)
2. Hand the brief to Claude Code (precise execution)
3. Review output with Gemini (catching edge cases)
4. Final polish back in Claude Code
Why: Each model has different strengths.
Claude Code orchestrates the whole thing
through skills and sub-agents.Stage 4 Checklist
This Is What a Complete System Looks Like
30+ agents, 50+ skills, 6+ hooks, harness gates, compound learning, progressive disclosure, agent coordination—all configured and working together. This system took months of iteration to build. The Second Brain delivers a customized version in hours.
See What's IncludedIntegration: MCP Connections
MCP (Model Context Protocol) is a standard way to connect Claude Code to external tools. Think of it as USB-C for AI - one protocol connects to everything.
┌─────────────────────────────────────────────┐
│ Claude Code │
│ └─ MCP Client (built-in) │
└─────────┬───────────────────────────────────┘
│
├─ MCP Server: Gmail
│ └─ Functions: send_email, search, read
│
├─ MCP Server: HubSpot
│ └─ Functions: create_contact, update_deal
│
├─ MCP Server: PostgreSQL
│ └─ Functions: query, insert, update
│
└─ MCP Server: Google Calendar
└─ Functions: create_event, list_eventsBusiness Tools
- • HubSpot (CRM)
- • Gmail
- • Google Calendar
- • Notion
- • Slack
Dev Tools
- • GitHub
- • Linear
- • PostgreSQL
- • Jira
Analytics
- • Google Analytics
- • Stripe
- • Mixpanel
Real Example: Sales Operations
Task: Tag 200 webinar attendees in HubSpot
Without MCP: 35 minutes
- 1. Export attendees from Zoom
- 2. Open HubSpot
- 3. Match emails manually
- 4. Bulk import with tags
- 5. Fix errors
With MCP: 3 minutes
- 1. "Tag attendees from webinar.csv"
- 2. Claude connects to HubSpot
- 3. Done.
Teaching Point: MCP setup can be the biggest time barrier
Andrew (product director, Netherlands) built automated job search tracking system—LinkedIn scraper → database → follow-up scheduler. DIY MCP setup: 6-8 hours of OAuth debugging. With Second Brain templates: 20 minutes.
See it in practice
One MCP, Three Platforms, 96 Seconds
Abstract diagrams are nice. Here's what an MCP actually does: a single command pulls 20 posts each from Reddit, X, and LinkedIn — 60 data points analyzed in under 2 minutes. No copy-pasting between tabs.
MCP OAuth Debugging Averages 6-8 Hours
The Done-With-You package includes a live configuration session where we connect your Gmail, Calendar, CRM, and database tools together—plus templates you can reuse for future tool connections.
See Done-With-YouStage 5 Checklist
What Customers Actually Built (First Month)
Week 1: Damian (AI Consultant, Switzerland)
Social media content automation system. Capture ideas → Extract LinkedIn voice DNA → Generate carousel PDFs → Schedule posts. "A friend just told me about this new super cool solution he bought. It was what I built in two days with the Second Brain."
Week 2: Andrew (Product Director, Netherlands)
Automated job search tracking system. LinkedIn scraper, application database, follow-up scheduler. Replaced a commercial SaaS tool he was paying monthly for—built in two days with the Second Brain.
Week 3: Dominic (Cloud Consultant)
Automated weekly client status reports. Salesforce MCP integration pulling deal data, Microsoft Graph for calendar/email context. Saved 3 hours/week on admin work.
Week 4: Richard (Management Consultant, Australia)
After struggling solo for a week: "You've solved for me on this call what I was still doing wrong—I was prompting it like it was a chatbot. Whereas by using commands and agents, you can effectively have that manage your daily workflow for you. Definitely worth the money to save me three to four months just hacking my way through it."
Next Steps: Your Path Forward
You've seen what a complete system looks like: progressive disclosure architecture, enforceable hooks, compound learning, the skills-vs-agents framework, MCP integrations. Building this from scratch takes 3-4 weeks of trial and error—debugging file structures, writing hooks, testing skill triggers, configuring MCP OAuth flows, designing memory schemas...
Or you can skip all that and get a customized version in hours.
"This was one of the best investments for me and my business. It opened up the world of easily configurable agentic work. The regular AI tools now just seem like child's play."
— Damian, AI Consultant • Done-With-You customer who built a complete content automation system in 2 days
The complete picture
5 Tasks. 20 Minutes. One System.
Email triage, customer responses, financial analysis, LinkedIn content, and a landing page — all kicked off from the same terminal in a single session. This is what the complete system delivers.
Choose Your Path
$197
DIY Package
Complete Second Brain repository with agents, commands, and skills. You handle setup yourself.
- ✅ Full Second Brain repository
- ✅ Pre-configured hooks system
- ✅ MCP configuration guides
- ✅ Documentation-based support
Timeline: 1-2 weeks setup
$597
Kickstart Package
Everything in DIY + AI agent configures your system from questionnaire answers.
- ✅ Everything in DIY, personalized for you
- ✅ AI agent builds system from your answers
- ✅ 48-hour priority async support
- ✅ First year MemoryOS Pro included ($349 value)
Timeline: Working in minutes, not weeks
$2,497
Done-With-You
Everything in Kickstart + 2-hour onboarding call with Iwo. Full infrastructure deployment.
- ✅ Everything in Kickstart
- ✅ 2-hour onboarding call with Iwo
- ✅ Full infrastructure deployment (Vercel + Supabase)
- ✅ Monthly Second Brain updates
- ✅ 12 months MemoryOS included
Timeline: Productive in 2 hours
Appendix: Advanced Patterns
Keep Learning
Coworking Session Recording
60-minute live session with 10+ demos — daily planning, content creation, MCP research, WhatsApp agents, and more.
AI Chief of Staff Template
Free agent template with full config, knowledge base structure, and customization guide for your role.
Official Documentation
These are the canonical sources. This tutorial adds the practical "how" and "when" that official docs don't cover.
Ready to Build Your Own Second Brain?
Three packages available - from DIY templates to done-with-you implementation.
See Packages30-day money-back guarantee • Save 10+ hours in first month or full refund