← All decks← Semua Deck
>_ run
● Course Summary · 3h 18m● Ringkasan Kursus · 3 Jam 18 Menit

Claude Code Advanced

How to use Claude Code at an advanced level — from system prompts to a fleet of agents.

Cara pakai Claude Code di level lanjutan — dari sistem prompt sampai fleet of agents.

Nick — Left Click CLAUDE.md Agent Harness Auto Research
Why Listen to HimKenapa Dengar Dia

Not a demo — this is what earns $4M a year.

Bukan demo, ini dipakai cari $4 juta/tahun.

The presenter uses Claude Code and AI agents in his own business every day to generate that profit, while also teaching ~2,000 people how to use these tools.

Presenter memakai Claude Code dan AI agents di bisnisnya sendiri setiap hari untuk menghasilkan profit itu, sekaligus mengajar ~2.000 orang cara memakai tools ini.

This course is not for beginners — it assumes you already have a basic Claude Code foundation.

Kursus ini bukan untuk pemula — mengasumsikan kamu sudah punya fondasi Claude Code dasar.

8 Topics Covered8 Topik yang Dibahas
  1. Advanced CLAUDE.md & system promptsCLAUDE.md & system prompt lanjutan
  2. Agent harness & large projectsAgent harness & project besar
  3. Agent teams & parallelizationAgent teams & parallelization
  4. Skills & sub-agentsSkills & sub-agents
  5. Auto research (Karpathy's approach)Auto research (pendekatan Karpathy)
  6. Browser automation & computer useBrowser automation & computer use
  7. Performance fluctuation & model alternativesPerformance fluctuation & alternatif model
  8. Workspace org, security, and the futureWorkspace org, security, masa depan
CLAUDE.md & System Prompts

CLAUDE.md is actually 4 things at onceitu sebenarnya 4 hal sekaligus

1 · Knowledge Compression

Compresses the entire workspace into one dense file — Claude doesn't need to read file after file every session.

Meringkas seluruh workspace jadi satu file padat — Claude tidak perlu baca file demi file tiap sesi.

2 · User Preferences

Personal conventions & preferences not yet natively supported by Anthropic — until they catch up.

Konvensi & preferensi personal yang belum di-native-support Anthropic — sampai mereka mengejar ketinggalan.

3 · Declaration of Capabilities

Explicitly states what Claude IS and ISN'T allowed to do — so it doesn't misjudge its own capability.

Menyatakan eksplisit apa yang BOLEH dan BISA dilakukan Claude — supaya dia tidak salah menaksir kemampuannya sendiri.

4 · Log of Failures & Successes

Notes on what has already failed/worked — so the same solution isn't retried from scratch.

Catatan apa yang sudah gagal/berhasil — supaya solusi yang sama tidak dicoba ulang dari nol.

Function 1 / 4Fungsi 1 / 4

Knowledge Compression

Instead of Claude opening every folder and file to understand the codebase, CLAUDE.md stores a super-dense summary: what it contains, why it was built, where the important parts are.

Daripada Claude membuka tiap folder dan file untuk memahami codebase, CLAUDE.md menyimpan ringkasan super-padat: apa isinya, kenapa dibuat, di mana letak yang penting.

The effect isn't just token savings — reasoning quality also improves because the context window used is far smaller and more relevant.

Efeknya bukan cuma hemat token — kualitas reasoning juga naik karena context window yang dipakai jauh lebih kecil dan lebih relevan.

Compression RatioRasio Kompresi
~45×

An 827-word file (~1,100 tokens) can be compressed into ~22 tokens in CLAUDE.md — enough for Claude to know "what this file is about" without opening it.

Satu file 827 kata (~1.100 token) bisa diringkas jadi ~22 token di CLAUDE.md — cukup untuk Claude tahu "file ini soal apa" tanpa membukanya.

Function 2 / 4Fungsi 2 / 4

User Preferences & ConventionsKonvensi

Output FormatFormat Output

"Always return absolute file paths" — so they can be clicked directly in any editor.

"Selalu kembalikan absolute file path" — supaya bisa langsung diklik di editor manapun.

Coding ConventionsKonvensi Coding

Preferred language/paradigm (e.g. functional style in Rust), standard project folder structure.

Bahasa/paradigma pilihan (mis. functional di Rust), struktur folder standar proyek.

How Unfamiliar Tools WorkCara Kerja Tool Asing

"When using an unfamiliar tool, read the API docs first — if the docs are JS-rendered, use Chrome DevTools MCP."

"Kalau pakai tool yang belum familiar, baca API docs dulu — kalau docs di-render JS, pakai Chrome DevTools MCP."

Anthropic keeps adding built-in preferences — but they always lag behind advanced users' needs. This is the gap CLAUDE.md fills.

Anthropic terus menambah preferensi bawaan — tapi selalu tertinggal dari kebutuhan advanced user. Ini celah yang diisi CLAUDE.md.

Function 3 / 4Fungsi 3 / 4

Declaration of Capabilities

Claude still often underestimates its own agentic capability if it isn't explicitly stated in the system prompt.

Claude masih sering underestimate kemampuan agentic-nya sendiri kalau tidak dinyatakan eksplisit di system prompt.

Real Story from the PresenterCerita Nyata dari Presenter

Asked "how long will X take?" — Claude answered "~3 months, because it has to build this and that."

Ditanya "berapa lama X akan selesai?" — Claude jawab "~3 bulan, karena harus bangun ini itu."

When the real answer was: "You're the one I'm asking to build it. You can finish in 5 seconds — just do it."

Padahal jawabannya: "Kamu yang saya minta bangun. Kamu bisa selesai dalam 5 detik — kerjakan saja."

Without an explicit declaration of what it is "allowed and able" to do itself, Claude treats itself like someone else who needs permission first.

Tanpa deklarasi eksplisit soal apa yang "boleh dan bisa" dilakukan sendiri, Claude memperlakukan dirinya seperti orang lain yang harus diminta izin.

Function 4 / 4Fungsi 4 / 4

Log of Failures & Successes

20% worth trying layak dicoba 80% of the solution space is already proven to fail 80% ruang solusi sudah terbukti gagal

Every file, every line of code that gets built successfully is hard-won — it already cost tokens and time. It's a waste if the lesson isn't passed on to the next session.

Tiap file, tiap baris kode yang berhasil dibangun itu hard-won — sudah menghabiskan token dan waktu. Sayang kalau pelajarannya tidak diwariskan ke sesi berikutnya.

The analogy: out of the entire "solution space" Claude might try, this log crosses out ~80% of options already proven to fail — leaving the 20% genuinely worth trying.

Analoginya: dari seluruh "ruang solusi" yang mungkin dicoba Claude, log ini mencoret ~80% opsi yang sudah terbukti gagal — menyisakan 20% yang benar-benar layak dicoba.

Two-Scope StructureStruktur Dua Scope

Global vs Local CLAUDE.md

Global — ~/.claude/CLAUDE.md

Always loaded, in every session

Selalu ter-load, di sesi manapun

High-level preferences, personal reasoning & communication style, agency capabilities, general token-saving strategy.

Preferensi tingkat tinggi, cara reasoning & komunikasi pribadi, agency capabilities, strategi hemat token secara umum.

Local — .claude/CLAUDE.md

Specific to one repository

Khusus satu repository

Low-level knowledge: project description, specific API docs, technical conventions only relevant here.

Pengetahuan level rendah: deskripsi project, API docs spesifik, konvensi teknis yang cuma relevan di sini.

Because global gets injected into EVERY session, keep it high-level — don't bloat it with details specific to one project.

Karena global di-inject di SETIAP sesi, taruh yang high-level di sana — jangan bikin bengkak dengan detail spesifik satu project.

🧪 Workshop

Try this in your own Claude Code setup this week

Coba ini di setup Claude Code kamu sendiri minggu ini

1
/init — knowledge compression: let Claude read the whole workspace and compress it into a local CLAUDE.md./init — knowledge compression: biarkan Claude baca seluruh workspace, ringkas jadi CLAUDE.md lokal.
2
Add preferences & conventions — project-specific rules: folder structure, deploy flow, where the API docs live.Tambah preferensi & konvensi — aturan spesifik project: struktur folder, flow deploy, lokasi API docs.
3
Run the meta-learning loop — after finishing a feature, ask: "how could this have been faster or cheaper in tokens?"Jalankan meta-learning loop — selesai fitur, tanya: "gimana caranya lebih cepat/hemat token?"
4
Write it to Lab Notes — save the answer as a running "don't do this again" log in CLAUDE.md.Simpan ke Lab Notes — catat jawabannya sebagai running log "jangan lakukan lagi" di CLAUDE.md.
Periodic Workflow — GlobalWorkflow Berkala — Global

Why the Review Must Be Manual

Kenapa Review-nya Harus Manual

After hundreds of runs, run /insights to analyze the full session history → a report of recurring patterns.

Setelah ratusan run, jalankan /insights untuk menganalisis semua histori sesi → laporan pola berulang.

But the manual review before something enters global CLAUDE.md is the one step that can't be fully automated — because global gets used across EVERY future project.

Tapi review manual sebelum masuk ke global CLAUDE.md itu satu-satunya langkah yang tidak boleh diotomasi penuh — karena global dipakai di SEMUA project masa depan.

Compounding Error MathMatematika Compounding Error

Claude 1: 90% accurate Claude 2: 90% accurate Claude 3: 90% accurate Total = 0.9³ = 73% (NOT 90%)

Claude 1: 90% akurat Claude 2: 90% akurat Claude 3: 90% akurat Total = 0.9³ = 73% (BUKAN 90%)

The more AI steps without a human in the loop, the more error accumulates — especially for a file that becomes the reference for every project.

Makin banyak langkah AI tanpa human-in-the-loop, makin besar error yang menumpuk — terutama untuk file yang jadi acuan semua project.

SynthesisSintesis

Local ↔ Global, Infinity Loop

LOCAL GLOBAL

The local loop runs per project — fast, granular, per-feature. After enough runs, the insight gets distilled up into the global loop used across all projects.

Local loop jalan tiap project — cepat, granular, per-fitur. Setelah cukup banyak run, insight-nya didistilasi naik ke Global loop yang dipakai lintas semua project.

Neither loop is ever "done" — it's a continuous improvement cycle, not a one-time setup.

Keduanya tidak pernah "selesai" — ini siklus perbaikan berkelanjutan, bukan setup sekali jalan.

Agent Harness

Everything That Wraps the LLM

Semua yang Membungkus LLM

SYSTEM PROMPT (CLAUDE.md) TOOLS (bash, MCP, browser…) LLM (Claude) galaxy-brain intelligence PARAMETERS (limit, auto-compact)

Definition: an agent harness is everything outside the model itself — the system prompt, tools, and parameters that wrap it.

Definisi: agent harness adalah segala sesuatu di luar model itu sendiri — system prompt, tools, dan parameter yang membungkusnya.

Without a harness, an LLM is just a text box. With one, it can control a computer, execute code, browse the web — and that's what makes Claude Code different from a plain chat model.

Tanpa harness, LLM cuma kotak teks. Dengan harness, ia bisa kontrol komputer, eksekusi kode, browsing — dan itulah yang membuat Claude Code beda dari sekadar chat model.

Harness EcosystemEkosistem Harness

Claude Code Isn't the Only Harness

Claude Code Bukan Satu-Satunya Harness

🟠

Claude Code

Anthropic's official harness — currently the most popular and most mature.

Harness resmi Anthropic — paling populer & paling matang saat ini.

🤖

Droid (Factory AI)

An open-source harness, an alternative outside the Anthropic ecosystem.

Harness open source, alternatif di luar ekosistem Anthropic.

🍐

pi.d.ai (Pear AI)

Open source & model-agnostic — plug in Claude, GPT, or any model.

Open source & model-agnostic — bisa pasang Claude, GPT, atau model apa pun.

Agent Teams & Extreme Parallelization

3 Reasons It Has to Be Parallel

3 Alasan Kenapa Harus Paralel

1 · Getting Slower1 · Makin Lambat

Autonomous agents used to take ~30 seconds/task; now it can be 15 minutes because far more gets done at once.

Autonomous agent dulu ~30 detik/task, sekarang bisa 15 menit karena makin banyak yang dikerjakan sekaligus.

2 · More Token-Efficient2 · Lebih Hemat Token

A shorter context window per agent = better reasoning performance.

Context window yang lebih pendek per-agent = performa reasoning yang lebih baik.

3 · Agents Are Stochastic3 · Agent itu Stochastic

Each run gives a different result — running many agents at once captures a wider variety of answers.

Tiap run hasilnya beda — menjalankan banyak agent sekaligus menangkap lebih banyak variasi jawaban.

Work PatternPola Kerja

Fan-Out → Fan-In → Synthesizer

┌── Research Agent 1 (Haiku) ──┐ │── Research Agent 2 (Haiku) ──┤ QUERY ──▶ Fan Out ┼── Research Agent 3 (Haiku) ──┼──▶ Fan In ──▶ Synthesizer ──▶ Answer │── Research Agent 4 (Haiku) ──┤ (Sonnet / Opus) └───────────────────────────────┘
┌── Research Agent 1 (Haiku) ──┐ │── Research Agent 2 (Haiku) ──┤ QUERY ──▶ Fan Out ┼── Research Agent 3 (Haiku) ──┼──▶ Fan In ──▶ Synthesizer ──▶ Jawaban │── Research Agent 4 (Haiku) ──┤ (Sonnet / Opus) └───────────────────────────────┘
Sequential
25 min25 menit

5 research + 5 + 5 + 5 + 5

Parallel
10 min10 menit

Research runs simultaneously, then synthesis

Research jalan bersamaan, lalu sintesis

Advanced TechniqueTeknik Lanjutan

Stochastic Consensus

Run the EXACT SAME agent 3-5 times. Because LLMs are stochastic, each run tends to surface slightly different ideas.

Jalankan agent yang PERSIS SAMA sebanyak 3-5 kali. Karena LLM bersifat stokastik, tiap run cenderung memunculkan ide yang sedikit berbeda.

Merge ALL the unique outputs — don't just pick the single "best" one.

Gabungkan SEMUA output unik — bukan cuma pilih satu yang "terbaik".

ExampleContoh

Run 1: [A, B, C] Run 2: [A, B, D, E] Run 3: [A, B, Z, Q] Unique: A,B,C,D,E,Z,Q

~2.5× more unique ideas compared to relying on a single run.~2.5× lebih banyak ide unik dibanding hanya mengandalkan satu kali run.

Auto Research — Karpathy's ApproachAuto Research — Pendekatan Karpathy

An Agent That Keeps Optimizing Itself

Agent yang Terus Mengoptimasi Diri Sendiri

Loop: 1. Claude builds a feature 2. Evaluate — "how could this be better?" 3. Implement the improvement 4. Measure a concrete metric 5. Repeat
Loop: 1. Claude bangun fitur 2. Evaluasi — "bagaimana ini bisa lebih baik?" 3. Implementasikan improvement 4. Ukur metrik konkret 5. Ulangi

This isn't a one-shot optimization — the longer the loop runs, the better the results, because each iteration prunes away the solution space that's already failed.

Ini bukan optimasi sekali jalan — makin lama loop-nya berjalan, makin baik hasilnya, karena tiap iterasi memangkas ruang solusi yang sudah gagal.

Presenter's Real ExampleContoh Nyata Presenter

Auto-Researcher Left Running for a Long Time

Auto-Researcher Dibiarkan Jalan Lama

Baseline Load TimeBaseline Load Time
1,802ms
After Running for a WhileSetelah Berjalan Lama
~8,000ms*

*The presenter shared this as evidence of "significant improvement after being left running for a while" — the point is the compounding impact of auto-research over time, not the absolute number.

*Presenter menyampaikan ini sebagai bukti "improvement signifikan setelah dibiarkan berjalan lama" — poin utamanya: dampak auto-research majemuk seiring waktu, bukan angka absolutnya.

Browser Automation

Computer Use vs Browser Use

Computer Use

Claude clicks, types, and navigates general desktop UI — used when the task spans native applications.

Claude mengklik, mengetik, dan menavigasi UI desktop secara umum — dipakai kalau tugasnya lintas aplikasi native.

Browser Use

Specifically for web-based tasks — faster & more precise because it's DOM-aware, used when the target is a web app.

Khusus untuk task berbasis web — lebih cepat & presisi karena DOM-aware, dipakai kalau targetnya web app.

Choose based on the target: native app → computer use, web app → browser use.

Pilih sesuai target: aplikasi native → computer use, aplikasi web → browser use.

Reliability

Performance Goes Up and Down, and That's Normal

Performa Naik-Turun, Itu Wajar

Claude Code's quality sometimes fluctuates day to day — this isn't a bug on your end.

Claude Code kadang berfluktuasi kualitasnya dari hari ke hari — ini bukan bug di sisi kamu.

Mitigation: Diversify ModelsMitigasi: Diversifikasi Model

Bundle multi-agent orchestration with other models (Codex, Gemini, etc.) alongside Claude.

Bundle multi-agent orchestration dengan model lain (Codex, Gemini, dll) di samping Claude.

The best solution is actually running several models at once — diversifying the risk of bias/weaknesses inherent to any single model.

Solusi terbaik justru menjalankan beberapa model sekaligus — mendiversifikasi risiko bias/kelemahan yang melekat pada satu model tertentu.

Work OrganizationOrganisasi Kerja

3 Workspace Levels

3 Level Workspace

👤

Personal

Personal experiments & tools.

Eksperimen & tools milik sendiri.

🏢

Business

Internal projects that support your own business operations.

Project internal yang menopang operasional bisnis sendiri.

🤝

Client ProjectsClient Projects

Work for clients — selling agentic engineering as a service.

Kerja untuk klien — menjual agentic engineering sebagai layanan (service).

SecurityKeamanan

Security Scaled to the Assets You Have

Security Sebesar Aset yang Kamu Punya

Auto Mode

Claude's newest feature that runs full autonomous mode without needing to flip permissions manually one by one.

Fitur terbaru Claude yang menjalankan mode otonom penuh tanpa perlu ubah permission manual satu-satu.

OAuth

Authentication considerations when connecting Claude to third-party services.

Pertimbangan autentikasi saat menghubungkan Claude ke layanan pihak ketiga.

Security Audit PromptSecurity Audit Prompt

A security-audit prompt (shared for free in his Vibe Coding course) — just run it once at the end of a project to catch the "low-hanging fruit."

Prompt audit keamanan (dibagikan gratis di kursus Vibe Coding-nya) — cukup dijalankan sekali di akhir project untuk menangkap "low hanging fruit".

The core idea: security risk scales with how much is at stake — it doesn't need excessive paranoia for a small project.

Intinya: risiko keamanan proporsional dengan seberapa besar aset yang kamu pertaruhkan — bukan sesuatu yang butuh paranoia berlebihan untuk project kecil.

The Future of Claude & Agentic EngineeringMasa Depan Claude & Agentic Engineering

4 Low-Risk Predictions

4 Prediksi Berisiko Rendah

Not random guesses — these are patterns the presenter considers very likely true, based on the trajectory already visible today.

Bukan tebakan acak — ini pola yang presenter anggap kemungkinan besar benar berdasarkan trajektori yang sudah terlihat sekarang.

Prediction #1Prediksi #1

Human Involvement Keeps Declining

Keterlibatan Manusia Terus Menurun

From the Wright Brothers flying the plane themselves, to modern aircraft engineers — autopilot has taken over most of the work.

Dari Wright Brothers yang menerbangkan pesawat sendiri, menjadi aircraft engineer modern — autopilot mengambil alih mayoritas pekerjaan.

Auto Mode in Claude is a microcosm of this trend.

Auto Mode di Claude adalah mikrokosmos dari tren ini.

Shifting RolePergeseran Peran

We'll all become a kind of "CEO" managing a fleet of agents — not a pilot flying the plane themselves.

Kita semua akan jadi semacam "CEO" yang mengelola fleet of agents — bukan pilot yang menerbangkan pesawat sendiri.

The ratio flips: instead of 100 people working on 1 task, it's 1 person managing 100 tasks at once. Leverage goes up drastically.

Rasio berubah: bukan 100 orang mengerjakan 1 tugas, tapi 1 orang mengelola 100 tugas sekaligus. Leverage naik drastis.

Prediction #2Prediksi #2

Software Quality Is No Longer a Moat

Kualitas Software Bukan Lagi Moat

AnalogyAnalogi

"You can recode Netflix in 5 minutes with 3-4 agents." If anyone can build it, what's Netflix's selling point?

"Kamu bisa coding ulang Netflix dalam 5 menit dengan 3-4 agent." Kalau siapa saja bisa membuatnya, apa nilai jual Netflix?

Moat Shifts ToMoat Bergeser Ke

Distribution, market reputation, and legal/compliance — no longer how good the code underneath is.

Distribusi, reputasi pasar, dan legal/compliance — bukan lagi seberapa bagus kode di baliknya.

The implication: the subscription SaaS model might not last much longer — people will choose between buying a finished product vs. "$19 + 30 minutes of tokens" to build it themselves.

Implikasinya: model SaaS berlangganan mungkin tidak lagi bertahan lama — orang akan memilih antara beli produk jadi vs "$19 + 30 menit token" bikin sendiri.

Prediction #3Prediksi #3

The Pace of Change Will Keep Accelerating

Laju Perubahan Akan Terus Berakselerasi

chimpsimpanse humanmanusia Einstein

The graph of "model intelligence over time" isn't linear — it's closer to exponential. Interestingly, the intelligence range from chimpanzee → average human → Einstein is actually quite close together if measured by neuron count.

Grafik "kecerdasan model dari waktu ke waktu" tidak linear — mendekati eksponensial. Menariknya: rentang kecerdasan simpanse → manusia rata-rata → Einstein sebenarnya berdekatan jika diukur dari jumlah neuron.

Current models sit around "chimpanzee-smart" level — but because that range is narrow, the jump to "galaxy-brain" could happen much faster than expected.

Model saat ini ada di sekitar level "chimpanzee-smart" — tapi karena rentangnya sempit, lompatan ke "galaxy-brain" bisa terjadi jauh lebih cepat dari perkiraan.

Prediction #4Prediksi #4

Early Adopters AreEarly Adopter adalah "The 1%"

Less than 1% of the world's population knows what an agent harness is — and far fewer than that can actually use one well. This creates a new productivity divide, not just a wealth divide.

Kurang dari 1% populasi dunia tahu apa itu agent harness — dan yang benar-benar bisa memakainya jauh lebih sedikit lagi. Ini menciptakan productivity divide baru, bukan cuma wealth divide.

"The future is here, it's just unevenly distributed."

— William Gibson

SummaryRangkuman

8 Building Blocks of Advanced Claude Code

8 Blok Bangunan Advanced Claude Code

📝

CLAUDE.md

4 functions + global/local + workflow loop

4 fungsi + global/local + workflow loop

🏗️

Agent Harness

System prompt + tools + LLM + parameters

System prompt + tools + LLM + parameter

Parallelization

Fan-out/fan-in + stochastic consensus

🔬

Auto Research

Self-optimization loop, Karpathy

Loop optimasi diri sendiri, Karpathy

🌐

Browser Automation

Computer use vs browser use

📉

Reliability

Diversify models across providers

Diversifikasi model lintas provider

🔒

Security

Auto mode, OAuth, audit prompt

🔮

The FutureMasa Depan

4 predictions, productivity divide

4 prediksi, productivity divide

You're no longer the pilot. You're the CEO of a fleet of agents.

Kamu bukan lagi pilot. Kamu jadi CEO dari fleet of agents.

A good CLAUDE.md, the right harness, and the courage to run many agents at once — that's the real currency of productivity in this era.

CLAUDE.md yang baik, harness yang tepat, dan keberanian menjalankan banyak agent sekaligus — itulah mata uang produktivitas yang sebenarnya di era ini.

DoneSelesai Source: Nick — Left ClickSumber: Nick — Left Click