Build & Sell with Claude Code — The 10-Hour Course Distilled

Claude Code isn’t just a coding assistant — it’s an autonomous agent platform that rewrites how software gets built. This post distills 10+ hours of course content into the key mental models, hacks, and business frameworks you actually need.

---

🧠 The Core Mental Model — What Claude Code Actually Is

---

Claude Code runs in your terminal and can read files, write code, run commands, browse the web, spin up sub-agents, and self-correct — all without you babysitting every step.

Think of it less like GitHub Copilot and more like a junior developer you can delegate entire projects to.

---

The Three Layers of Claude Code

Layer	What It Does
CLAUDE.md	Your permanent system prompt — injected at the start of every session
Skills / Tools	Reusable prompt files and MCP servers that extend capabilities
Sub-agents / Teams	Parallel workers with isolated context that report back to the main session

---

How a Session Actually Works

flowchart LR
    A([You]) --> B[Main Claude Code Session]
    B -->|delegates| C[Sub-agent 1\nResearch]
    B -->|delegates| D[Sub-agent 2\nTesting]
    B -->|delegates| E[Sub-agent 3\nCode Review]
    C --> F([Main session\nreceives report])
    D --> F
    E --> F
    style B fill:#4A90D9,color:#fff
    style C fill:#5BA85A,color:#fff
    style D fill:#E8A838,color:#fff
    style E fill:#9B6EBD,color:#fff
    style F fill:#4A90D9,color:#fff

---

⚙️ Context Management — The #1 Thing Beginners Get Wrong

---

Context rot is real. When your context window fills up, output quality degrades — and if you’re running 10+ sessions in parallel, you often won’t notice until it’s too late.

---

The Context Hygiene Playbook

1. Keep CLAUDE.md lean — 100–200 lines max. Every line is injected into every session. Route non-essential docs to external files and just reference them.
2. Use /compact at 60% — Compress history mid-session, keeping only critical decisions. You can tell it what to keep: /compact but keep all API integration decisions.
3. Use /clear between tasks — Wipe history when switching topics. Your CLAUDE.md and files persist; your conversation doesn’t.
4. Run /context — See exactly what’s eating your tokens. MCP servers can be surprisingly expensive — sometimes a direct API endpoint costs far fewer tokens.
5. Max 3–4 parallel sessions — More than that and you lose track of what’s going on. Quality over quantity.

---

Sub-agents for Heavy Lifting

If you have a 15-minute YouTube transcript to analyse, don’t let your main session read it. Delegate to a sub-agent: “read this, give me the key highlights.” Keep the main thread clean.

flowchart TD
    A[Main Session\n~Clean context] -->|"Read 200-page PDF\nSummarise findings"| B[Sub-agent\nHaiku model\nCheap + fast]
    B -->|"3-bullet summary"| A
    style A fill:#4A90D9,color:#fff
    style B fill:#5BA85A,color:#fff

Use Haiku for sub-agents doing heavy data processing (scraping, summarising, reading large files). Reserve Sonnet/Opus for the main thread where quality matters.

---

🔧 Git, GitHub & Work Trees — Safe Parallel Development

---

Git Concepts in 60 Seconds

Concept	Analogy
Repository	Bank account — stores every transaction (commit) ever made
Commit	Checkpoint — snapshot of the project at a point in time
Branch	Draft mode — make changes without touching the published version
Push / Pull	Sync to / from cloud
Merge	Combine a branch back into main once tested

---

Git Work Trees — Run Multiple Sessions Safely

The problem: normally you can only work on one branch at a time. Switch branches and you risk losing in-progress work.

The solution: work trees give each Claude Code session its own isolated branch and folder — no risk of agents overwriting each other.

# Each terminal gets its own workspace + branch
claude --worktree feature-login
claude --worktree bugfix-api

Claude Code has native work tree support built in. You can run 4–5 parallel sessions on the same project and merge them all back to main cleanly when done.

---

🤖 Browser Automation — When There’s No API

---

Claude Code can launch a real browser, fill out forms, click buttons, scrape pages, and run QA loops — even on sites where you’re already logged in.

---

The QA Self-Checking Loop

flowchart LR
    A[Build feature] --> B[Take screenshot]
    B --> C{Looks right?}
    C -->|No| D[Identify bugs\nfrom screenshot]
    D --> E[Fix code]
    E --> A
    C -->|Yes| F[Open Chrome DevTools\nCheck for JS errors]
    F -->|Errors found| D
    F -->|Clean| G[✅ Ship]
    style A fill:#4A90D9,color:#fff
    style G fill:#5BA85A,color:#fff
    style D fill:#D9534F,color:#fff

Browser automation is iterative — expect 4–5 loops before a script runs cleanly. That’s normal, not a failure. Each run, the script gets smarter.

---

Persistent Login Sessions

Claude Code can use your existing Chrome profile, meaning it can interact with platforms where you’re already logged in — Notion, CRMs, school communities, etc. First run requires manual login; subsequent runs use the saved session.

---

🛠️ The 32 Claude Code Hacks — Ranked by Impact

---

Beginner Hacks (1–10)

#	Hack	Why It Matters
1	/init on every project	Auto-generates CLAUDE.md from your codebase
2	Status line	See model, context %, and cost at a glance
3	Voice input (/v)	Talk to your terminal — great for long prompts
4	Keep context small	Focused context = better outputs, no exceptions
5	/context to find bloat	Diagnose token waste before it degrades quality
6	/compact at 60%	Compress mid-session, keep critical decisions
7	Always start in plan mode	Claude outlines steps before writing a line of code
8	Give problems, not commands	“How should we handle growth tracking?” beats “Write me a function”
9	Make Claude ask questions	“Ask me questions until you’re 95% confident you understand the task”
10	Self-checking to-do lists	Bake verification steps (screenshot, DevTools check) directly into the task list

---

Intermediate Hacks (11–22)

#	Hack	Why It Matters
11	Sub-agents for parallel work	Team of specialists > one generalist
12	Custom skills	Reusable .md prompt files in ~/.claude/skills/ — commit to GitHub for team use
13	Haiku for sub-agents	10× cheaper for bulk processing tasks
14	Refresh CLAUDE.md constantly	Log new patterns, gotchas, conventions after each session
15	CLAUDE.md routes to other files	Keep it lean; point to style guides and reference docs instead of embedding them
16	Exit early and reask	Don’t let Claude burn tokens going the wrong direction — hit Escape and correct
17	Challenge outputs aggressively	“Scrap that, try again with a completely different approach” often yields dramatically better results
18	/re for quick undos	Roll back to a previous conversation point without starting over
19	Hooks for notifications	Get a sound alert when a long-running session finishes
20	Feed screenshots	Claude can see — use this for design checks, error message analysis, and visual QA
21	Chrome DevTools	Test app functionality, not just appearance
22	Clone inspiration sites	Screenshot a design you love, feed it in, say “make it look like this”

---

Advanced Hacks (23–32)

#	Hack	Why It Matters
23	Parallel sessions + work trees	Run 5 isolated branches simultaneously, merge when done
24	API endpoints over MCP servers	If you only need one endpoint, loading a full MCP server wastes tokens
25	/loop for recurring tasks	Monitor deployments, PRs, error logs — runs in background, interrupts only when needed
26	Host on a VPS	Always-on sessions, SSH in from anywhere, control via Telegram
27	Remote control from phone	Start a build at your desk, steer it from your pocket
28	No-SQL data analytics	Connect BigQuery CLI → ask questions in plain English → Claude writes and runs the query
29	ultrathink	Allocates 32K thinking tokens for hard architecture decisions or when stuck after 2+ attempts
30	Fine-grained permissions	Explicitly allow safe commands + deny destructive ones (deny takes priority) = full speed, no danger
31	Agent teams	Agents that share a task list and can talk to each other — more cohesive than sub-agents for complex projects
32	Context7 MCP	Injects current library docs (Next.js, React, MongoDB…) before Claude writes code — eliminates hallucinated deprecated APIs

---

💼 Building a Business with Claude Code

---

The Core Mindset Shift

Stop selling AI agents. Start selling outcomes tied to real business pain.

Businesses don’t care if you used Claude Code, n8n, or duct tape. They care about time saved, money saved, and fewer errors.

---

The Diagnose → Solve → Value → Price Framework

flowchart LR
    A[🔍 Diagnose\nFind the pain point] --> B[🔧 Solve\nBuild the system]
    B --> C[💰 Value\nTranslate to numbers]
    C --> D[🏷️ Price\nAnchor to ROI]
    style A fill:#4A90D9,color:#fff
    style B fill:#5BA85A,color:#fff
    style C fill:#E8A838,color:#fff
    style D fill:#9B6EBD,color:#fff

Example: Client onboarding automation

Step	Action
Diagnose	Team spends 5 hrs/week onboarding clients
Solve	Automate 80% of that process
Value	200+ hrs/year saved × $50/hr = $10,000/year back
Price	“I’m saving you $10K/year — I’m charging $3K” → obvious yes

---

How to Get Your First Clients (No Audience Needed)

Three methods, in order of effectiveness when starting out:

1. Warm Outreach (Start Here)

Build a “trust map” — list 20 people you already know who run businesses. Reach out not to pitch, but to learn:

“I’m mapping the top time drains in your industry. Can I ask you 10 minutes of questions? No pitch.”

2. Referrals

After delivering value, ask the golden question: “Do you know any other business owners who might benefit from this?”

Only 11% of salespeople ask for referrals — yet 91% of happy clients would give one if asked. Ask every time.

3. The Trojan Horse Method (Highest Leverage)

Partner with agencies, consultants, and coaches who already have client relationships. Offer:

• Free AI discovery calls for their clients
• 20% rev share on any projects you close
• They look like heroes for bringing in an expert; you get warm introductions

Partner-sourced deals close 46% faster than cold outreach. You borrow trust instead of building it from scratch.

---

Value-Based Pricing — The Simple Math

Annual cost of manual process = (hours/week × 52) × hourly_rate
Your target price ≈ 10–15% of annual savings

Real example:

• 20 leads/week, 1 hr each to qualify → 20 hrs/week × $40/hr = $38,400/year
• Automated 80% → $30,720 saved/year
• Project price: 15% × $38,400 = $5,760 → they see 5× ROI in year 1 alone

---

The PRICE Framework

Letter	Step
P — Prepare	Ground yourself in value-based pricing, not hours
R — Research	Map the full manual workflow end-to-end
I — Identify	Calculate ROI, annual savings, opportunity cost
C — Communicate	Paint the “after” picture before saying a number
E — Expand	After delivery: maintenance, optimisation, V2 retainer

---

🚀 Delivery & Handover — The Professional Way

---

The Three Hosting Options

Option	Best For	Compliance
Client hosts	Almost all client work	✅ Cleanest — each client owns their instance
You host internally	Your own agency automations	✅ Fine as long as clients don’t log in
You host as product	SaaS / subscription platform	⚠️ Requires commercial license

Never run multiple clients on a single shared instance without a commercial license. This violates most automation platform terms of service.

---

The QA Loop Before Handover

flowchart TD
    A[Internal QA\nTest with real sample data] --> B{Pass?}
    B -->|No| C[Fix bugs\nUpdate prompts]
    C --> A
    B -->|Yes| D[Client QA\nSimple UI / form to test]
    D --> E{Client happy?}
    E -->|No| F[Prompt tuning\nMinor tweaks]
    F --> D
    E -->|Yes| G[Push to Production]
    style G fill:#5BA85A,color:#fff
    style C fill:#D9534F,color:#fff
    style F fill:#E8A838,color:#fff

---

Scope Creep — Protect Yourself

Define the exact list of functionality in the contract. When clients ask for more during the build:

“Love that idea — I’ll add it to the V2 backlog. For this version, we’re locking scope so I can deliver on time.”

This plants the seed for future work and protects your margins.

---

💡 One-Sentence Intuition

Claude Code turns you into a manager, not a coder — your job is to give clear goals, steer agents when they drift, and stack proof until you’re the AI partner businesses can’t afford to lose.

---

🔗 Related Resources

• Claude Code Docs
• Context7 MCP — live library documentation injection
• Pixel Agents VS Code Extension — visualise your agents as pixel art characters (Windows only for now)

---

---

Part of the AI Tools & Automation series. Next: running Claude Code on a VPS for always-on autonomous agents.