Vibe Off, Spec On — Building Real Offensive Tools with Claude Code
3-day in-person
Most "AI for hackers" courses are closed-lab cosplay — a curated exercise on rails wrapped around prompt screenshots, the same pre-made tool every time. This is the opposite. On Day 1 morning the class "votes on a real tool" to build over the next three days, writes a real PRD with Claude in the room, and ships it by Day 3.
Nobody — instructor included — knows what we'll land on. If you have experience with AI, likely, different students will ship very slightly different tools or use a different approach or architecture. That’s point and that’s what makes this class a ‘real life’ training.
At the end, the student will master Claude Code and discern from the Vibe code concept to really empower AI to build working, documented and well-designed tools.
The training foresees two parallel tracks every session:
a method track (e.g. Claude Code internals, custom skills, subagents, hooks, MCP authoring, eval harnesses, deployment, code review, distribution)
and a live build track on the chosen tool. Spec-driven with OpenSpec. No vibe coding. No premade labs. No "build me a million-dollar app" prompts.
The training teaches the same approach, developed in almost two years of R&D, used to build tools like FOIL (https://foil.peachstudio.be), a security code scanner. Foil surfaced multiple 0-days including a ~22-year-old zero-days — and reframed against Anthropic's Mythos Preview (Project Glasswing). Foil uses GPU inference, a custom LoRA trained model picturing Claude Code’s potential.
Course Overview
Two tracks running in parallel: a method/tool track (formal training) and a real-life build track on the class-voted tool (live driven by the instructor).
Day 1 — Tool, environment, and what we're building
S1 · Meet your new best friend (or girlfriend, considering the cost) — Claude Code 101 for offensive engineers.
Anatomy of the loop: tools, plan mode, sub-agents, MCP at a glance
Settings hierarchy, permissions, context window economics
Where Claude shines, where it lies, what the bill looks like at the end of a real R&D week
S2 · Make new Claude Code gurus — the official Anthropic toolkit.
It’s a deep-dive section. By the end of S2 a student can install, customize, extend, and distribute Claude Code as a hardened personal toolkit:
`CLAUDE.md` hierarchy (project / user / nested) and context augmentation done right
Custom skills — authoring, packaging, distributing
Subagents — Explore, Plan, security-reviewer, case study: the open sourced threat profiler skills (available on the Claude Skills Marketplace of PeachStudio)
Hooks — `UserPromptSubmit`, `PreToolUse`, `Stop`, `SessionStart`; using hooks as guardrails on offensive tooling
Custom slash commands
MCP servers — consuming and authoring your own
Output styles, status lines, keybindings
Background tasks, scheduled agents, plan mode, work trees
The plugins ecosystem — install, audit, ship your own
S3 · Vibe coding? No thanks.
Why offensive tooling must not be vibe-coded
Spec-driven development with OpenSpec: from idea to PRD, to architecture, and finally to code (testing included that Claude will take care of)
Patterns that survive when the agent goes sideways. Claude will always tell you ‘good catch’ but how to productively challenge to get better results?
Cross-cutting (Day 1): the class polls and votes on the tool to build over the next three days. PRD drafted live with Claude, human in the loop. Repo bootstrapped end of day.
Day 2 — Build, test, integrate
S4 · Testing and debug harness
TDD with Claude Code on real code
Eval-driven testing
Golden tests with Playwright and Claude Code Chrome Extension
Documentation and memory management
S5 · Deploy and integrate.
SaaS deployment patterns for offensive tooling
MCP for live-environment interaction
Dev inside containers, sandboxing, blast-radius control
Thinking smart, simplify your life
Integrate features and adding them to the PRD as developing
Cross-cutting (Day 2): the voted tool advances (with various LAB live sessions, on the second day the track 2 is predominant) — PRD → architecture → implementation → first integration tests for the offensive security tool chosen.
Day 3 — Distribute, secure, survive
S6 · Distribute your work.
Packaging and releasing offensive tooling safely, let's containerize it
Plugins, SDK, internal-only distribution patterns
Any additional requirement (e.g. Mobile apps on Apple stack? Ohh..)
Publishing on NPM and PyPi, and why not; brew –cask for quick install
Deploying our tool on Vercel or Railway
S7 · Local LLM with Claude (brief).
When sensitive ops can't leave the host
Understanding local AI restrictions: the real life beyond the fuzz and lies of running LLMs locally with Ollama or LMstudio
HuggingFace, how to choose a model that works, maybe, in Claude
Use of the opensource project claude-code-local (peachstudio)
Local-model trade-offs, hybrid setups, when to stay on Claude
Connecting a DGX-Spark
S8 · Secure your code, code review with Claude — before and after the Mythos era (yeahh because many tells code developed with AI is insecure)
Claude-driven code review workflows
What the Anthropic Mythos Preview (Project Glasswing, 2026-04-07) changed: autonomous zero-day discovery against OpenBSD (27-year-old TCP SACK bug) 🡪 lesson learned we will have by the time of delivery of this class.
Finding vulnerabilities with Foil (community edition), triage with Claude, ohh yes what a sweet spot: How I found a 22y old vuln on a popular bootloader/embedded firmware. CVE in attribution at the time of writing so, cannot disclose more.
What Mythos means for the tooling we just built — both as defenders and as offensive engineers
Cross-cutting close (Day 3): ship and demo the class-built tool. Retrospective on what worked, where Claude failed, what the audience would do differently and finally Beers to celebrate.
## What students will take home
Battle-tested Claude Code workflows for offensive R&D and developing of tools and small applications.
The full power-user toolkit: custom skills, subagents, hooks, slash commands, MCP servers, plugins, output styles — authored and distributed, not just consumed
Spec-driven discipline (OpenSpec) that prevents catastrophic vibe-coding in security tooling
An eval/test harness pattern for agentic and non-deterministic components
MCP integration patterns for real, live environments
A real, working tool the cohort designed and shipped together — optionally open-sourced
The trainer's two years of accumulated tweaks, dead-ends, and secrets — including the workflow behind FOIL (https://foil.peachstudio.be)
Target Audience
Developers, security researchers, and tool authors who want to stop watching demos and start shipping.
You don't need to be a hacker — you need to want to build something out of your ideas and creativity.
Pre-requisites
Comfortable with shell, git, and either Python or TypeScript
Working knowledge of offensive security fundamentals
Laptop with Docker installed
Active Claude Code account (Pro/Max or API credit).
A budget guidance note is sent two weeks before the training, so nobody hits a credit ceiling on Day 2. All exercises run on each student's own laptop. There is no shared lab infrastructure — by design.
Special equipment / licensing requirements
Student-side: laptop, Docker, Claude Code account (own subscription or API key with budget)
No room-side requirements beyond standard projector + whiteboard
All software used in the training is either open-source, free-tier, or covered by the student's own Claude Code subscription
Tainer Bio
Vito Rallo is a cybersecurity expert and product builder. Founder of Peach Studio (Brussels), an AI product studio at the intersection of security and generative systems. With 25+ years in technology and 18+ in cybersecurity, Vito previously led Red Team global units at Kroll and PwC and started his career as an ethical hacker on the IBM X-Force team. He co-founded Crimson7 in Brussels.
Today he works as Consult Partner for Cybersecurity in Kyndryl and free time he focuses on hands-on AI-assisted offensive R&D with Claude Code — including FOIL, a tool he developed largely with Claude that surfaced multiple ~22-year-old zero-days — and on training the next generation of practitioners to build, ship, and secure their own offensive tooling.
Vito is actively involved in cybersecurity research and innovation, and enjoys delivering keynotes and talks to security conferences.
