Agentic Coding Workshop

Quick Start

# Install Claude Code
npm install -g @anthropic-ai/claude-code

# Start (with full autonomy)
claude --dangerously-skip-permissions

Without --dangerously-skip-permissions you'll be confirming all day.

Core Principles

Monolithic > Multi-Agent — One agent, one task, one loop
Plan First — Explore first, then plan, then build
Small Tasks — Limit scope, restart on drift
Validate, Don't Hope — Tests, linter, build as feedback loop
Git as Safety Net — Checkpoints before big changes

The Ralph Loop

┌─────────────────────────────────────┐
│  1. Explore (understand codebase)   │
│  2. Plan (IMPLEMENTATION_PLAN.md)   │
│  3. Build (one task per iteration)  │
│  4. Error? → Back to 2              │
└─────────────────────────────────────┘

The plan persists between iterations. Each iteration starts fresh, reads the current state, and works on the next task.

You Are the Puppet Master

You do

High-level architecture
Task breakdown
Tech stack decisions
Final QA (like a PM)
You know the goal

Claude does

All the grunt work
Implementation
Writing tests
Debugging
Self-validation

CLAUDE.md — The Heart of It

Placement

File	Scope
`./CLAUDE.md`	Project root, commit to Git
`./CLAUDE.local.md`	Personal, add to .gitignore
`~/.claude/CLAUDE.md`	Global for all sessions
`./subdir/CLAUDE.md`	Monorepo, loaded on-demand

What Goes In

Commands — Build, Test, Lint with descriptions
Code Style — "ES modules, not CommonJS"
Conventions — Branch naming, commit style
Architecture — Core files, important utilities
Gotchas — Project-specific warnings

Control Behavior

Interview Mode: "For new features, ALWAYS clarify requirements before writing code."

Self-Evaluation: "Before completion: run tests, check build, fix errors independently."

Phase Separation: Research → Planning → Implementation (not all at once!)

Emphasis Keywords

Improve instruction-following:

IMPORTANT: Always run tests before committing
YOU MUST: Use TypeScript strict mode
NEVER: Commit directly to main
ALWAYS: Ask before architectural changes

Template

# Project Name

## Commands
- `make dev` — Start dev server
- `make test` — Run tests (ALWAYS before commit!)
- `make lint` — Check code style

## Conventions
- TypeScript strict, no any types
- Conventional Commits (feat:, fix:, ...)
- Branch: feature/*, fix/*

## IMPORTANT
- NEVER commit without tests
- ASK before architectural changes
- New feature? Clarify requirements first!

Golden Rule: Treat it like a prompt — test and refine iteratively. Press # to automatically add instructions.

Context Management

New session at ~40% context usage
Keep sessions focused
CLAUDE.md holds persistent context
Quality drops when context fills up

Red Flags — When Output is Bad

Claude starts building weird test scripts
Excessive mocking in tests
Claude is going in circles
More workarounds instead of real fixes

Solution: Don't patch. Context is polluted → New session, regenerate plan.

Voice Input

VoiceInk (Mac) + Parakeet model — faster than Whisper, runs locally.

Perfect for talking through plans
Give lots of context via voice
Claude creates first draft, you review

Not Just for Code

Workout tracking — Reports to coach
Podcast discovery — Follow guests
Article research — "Didn't I read something about X?"
Bookkeeping — Accountant questions

Health — Blood work + book context
YouTube — Summarize videos
Voice messages — Transcription + summary

Skills make the difference: Claude has access to your data.

Costs

What	Cost
Claude Code (Max 20x)	~€200/month
OpenRouter (Chat, Workflows)	~$20 lasts 1-2 months

Agentic Coding Workshop — Handout