Vibe Coding Workshop · Build your dream startup in a day · EBS Tallinn · May 9

Build a Team
With Claude Code.

Stop prompting. Start delegating.

claude — agent-team
scroll
The pitch

Lovable lets one of you be a 1-person startup.

Claude Code lets one of you be a 5-person startup.

Same Saturday. Five times the surface area.

The ceiling

Anthropic's researcher walked away.

A C compiler in Rust. From scratch. Builds the Linux kernel. He mostly watched.

16
agents working in parallel
100K
lines of Rust for a C compiler
human required (mostly)

You don't need $20K. You need 3 agents and 3 hours.

Source: Nicholas Carlini, Anthropic — "Building a C compiler with a team of parallel Claudes" · Feb 2026

Mental model

Three things people confuse.

Main agent
The session you talk to.
One conversation, one context window. Default Claude Code.
Subagent
A specialist you delegate to.
Fresh context. Does the work. Reports back. One-way street.
Agent team
A lead + named teammates.
Shared task list. Talk to each other. Self-coordinate.

↑ today's focus

The decision rule

Memorize this.

Subagent = go do this and tell me.

Team = figure it out together.

Quick errand → subagent. Real coordination → team.

The wow moment

A subagent is a markdown file.

That's the whole secret. Save it to .claude/agents/ and it's live.

code-reviewer.md
--- name: code-reviewer description: Senior code reviewer. Use PROACTIVELY after code changes for security, bugs, and bad patterns. tools: Read, Grep, Glob, Bash model: sonnet --- You are a senior reviewer with 15 years of experience. When invoked: 1. Run `git diff` to see what changed. 2. Focus only on changed code paths. 3. Return: 🔴 must-fix, 🟡 should-fix, 🟢 nice-to-have. 4. Cite line numbers. Suggest concrete fixes. Be direct. No padding. No restating the code back to me.
Live demo

Spawn a founder team.

One prompt. Three teammates. They coordinate themselves.

› claude
Create an agent team to build a TODO app with auth. Spawn three teammates: one for the backend API, one for the React frontend, one for tests. Coordinate through the shared task list. Quality bar: production-ready code, no shortcuts.
What just happened

Anatomy of a team.

LEAD Backend Frontend Tests they message each other directly
  • Each teammate has its own context window — no clutter.
  • Shared task list at ~/.claude/tasks/{team-name}/
  • You can pop into any pane and talk to them directly.
Steal these

Three teams to spawn today.

Copy. Paste. Spawn.

🛠
Build squad
MEMBERS
architect · frontend · backend · tester
BEST FOR
New features end-to-end.
› build-squad.prompt
Create an agent team called build-squad. Spawn four teammates: – architect: designs the data model and API contract first – frontend: implements UI components against the contract – backend: implements endpoints against the contract – tester: writes integration tests and runs them The architect produces the contract first; frontend and backend work in parallel from it; tester verifies last. Quality bar: production-ready, typed, tested.
🚀
Ship squad
MEMBERS
release-captain · changelog · deploy-checker
BEST FOR
Pre-deploy QA passes.
› ship-squad.prompt
Create an agent team called ship-squad to QA my app before I deploy. It's running at http://localhost:3000. Spawn three teammates: – release-captain: orchestrates and produces the final report – changelog-writer: drafts release notes from git log – deploy-checker: hits every route, checks console errors, flags broken images and 404s Each teammate uses Sonnet. Return one consolidated report.
🔍
Recon squad
MEMBERS
scout · synthesizer · skeptic
BEST FOR
Exploring an unfamiliar codebase or idea.
› recon-squad.prompt
Create an agent team called recon-squad to help me understand this codebase before I make changes. Spawn three teammates: – scout: maps the directory structure and entry points – synthesizer: explains how data flows end-to-end – skeptic: lists risky areas and likely bugs Each runs in read-only mode. Return one consolidated brief.
Templates

Two more ready-made subagents.

Drop into .claude/agents/ and they're live.

planner-researcher.md
--- name: planner-researcher description: Plans implementation before any code is written. Use PROACTIVELY before any feature or refactor of more than ~50 lines. tools: Read, Grep, Glob model: opus --- You are a thoughtful planner. When invoked: 1. Read enough of the codebase to understand the area being changed. 2. Produce a plan in this exact format: ## Goal One sentence. ## Approach 3–6 numbered steps. ## Files to touch List with one-line reason each. ## Risks / unknowns Bullet list. Be honest. ## Out of scope Bullet list. What we're NOT doing. Do not write code. Do not modify files. Plan only.
launch-copywriter.md
--- name: launch-copywriter description: Writes launch copy — landing page, tweet threads, emails — in the voice of the product. Use when shipping or relaunching. tools: Read, Grep model: sonnet --- You are a sharp launch copywriter. When invoked: 1. Read the README and any product docs to learn the product's voice. 2. Produce these three deliverables: – Hero: one-sentence value prop + 3 supporting bullets. – Tweet thread: 5–7 tweets, no emojis unless the product uses them. – Email: 80–120 words, one CTA. 3. Match the product's existing tone. Don't default to corporate. Never use these words: "revolutionary", "game-changer", "seamless", "leverage", "unlock", "supercharge".
Project memory

The CLAUDE.md every project should have.

Save in your project root. Every Claude Code session reads it automatically — including every teammate in a team.

CLAUDE.md
# Project Context Brief sentence about what this project does and who uses it. ## Stack - Frontend: [framework, language] - Backend: [framework, language] - Database: [Postgres, Mongo, etc.] - Hosting: [Vercel, Railway, etc.] ## Conventions - Code style: [Prettier? Biome? Black? Standard?] - Naming: [camelCase, snake_case, kebab-case where] - File layout: [where do new components go? new API routes go?] ## Quality bar - All new code is typed. - All new endpoints have at least one happy-path test. - No console.logs, no `any`, no commented-out code in PRs. ## When delegating to subagents or teams - Use `planner-researcher` before any feature larger than ~50 lines. - Use `code-reviewer` after any change before declaring done. - For features touching API + UI + tests, spawn a `build-squad` team. ## Things to NEVER do - Don't use placeholder data — read the seeded fixtures. - Don't invent file paths — verify with Glob first. - Don't refactor unrelated code while implementing a feature.
The things that burn hours

Six gotchas.

1
3–4 teammates is the sweet spot.
More is overkill. Tokens scale linearly.
2
Plan first. Spawn second.
No plan = agents wandering off. Use plan mode.
3
Teammates don't inherit your chat.
Tell the lead to embed context into tasks.
4
/clear between unrelated tasks.
Don't pollute context with old threads.
5
Tell the lead the quality bar.
It trickles down to every teammate.
6
Tokens are real. ~2–4× a single session.
Worth it for parallelism, not for one-line fixes.
Get started

Three steps. Two minutes.

1 · Install — pick your platform
macOS: brew install --cask claude-code
Linux: curl -fsSL https://claude.ai/install.sh | bash
Windows: irm https://claude.ai/install.ps1 | iex
2 · Run
cd your-project && claude
3 · Spawn your first agent
/agents

Send me a DM with what you build today — I want to see it.

QR code linking to Ahmed Aleryani's LinkedIn profile
Scan → connect on LinkedIn
Or just tap the link below
linkedin.com/in/ahmed-aleryani