Vibe Off, Spec On — Building Offensive Security Tools with Claude Code
A non-deterministic, hands-on training. The class votes on the tool. We build it together, without forgetting the theory.
AI is non-deterministic, so a training course can't offer labs with deterministic answers. Most "AI development" courses talk about vibe coding and hand you a closed, packaged exercise. This one doesn't. At the end of day one the class votes on a real offensive tool to build, writes a real PRD together with Claude, and uses it to release the tool, or a prototype, by the end of the course. Nobody, trainer included, knows in advance what the class will choose. Every session runs two tracks: a Method Track for theory and method, and a Live Build Track where each technique is applied straight away to the class's tool. The goal isn't just learning Claude Code: it's learning how to think when building an offensive tool, not an enterprise app.
What you'll walk away with
Two years of research and development on building with Claude Code — the methodology and the approach are what count.
Spec-driven, not vibe-coded
Real PRDs and spec-driven development with OpenSpec, combined with loop & harness engineering. Concepts that apply to any agentic coding harness, not only Claude Code.
Skills, tools-as-code & plugins
Custom skills, tools-as-code, subagents, hooks, MCP servers and plugins — written and distributed as open source.
Testing & debugging with AI
An AI-automated eval and test-harness pattern for agentic components, SaaS applications and Python tools; debug loops with /goal and Playwright.
Secure it at generation time
Shift-left secure-coding rules mapped to the OWASP Top 10, Claude-driven code review, and vulnerability hunting with FOIL and other skill-driven tools.
Sovereignty & local LLMs
When code can't land on a cloud service: open-weight models, local hardware from RTX to DGX Spark, and when to stay on Claude.
A first working PoC
Learn from the trainer's experience and from the rest of the class, and leave with a first proof of concept you helped build.
Three days, nine sessions
At least three intense days, extendable up to five. Everything runs on your own laptop; skills and tools are provided and released as open source.
The tool, the environment & what we'll build
- S1 · Claude Code 101 for offensive tool builders — the agent story, anatomy of the loop, settings and permissions, context management, memory and sessions, Context7, and what an R&D week really costs.
- S2 · The full power-user toolkit —
CLAUDE.md, skills and Record-a-Skill, tools-as-code, subagents and agent teams, hooks as guardrails, MCP repositioned, the plugin ecosystem. - Cross-cutting — the class proposes and votes on the tool, and closes the day with a first PRD draft.
Loop, build, test, integrate
- S3 · Spec-driven development, no vibe coding — from idea to PRD, architecture and code with OpenSpec.
- S4 · Loop & harness engineering — context → harness → loop;
/goaland/loop; what offensive security really needs. - S5 · Verification, testing & debug — the debug harness, Playwright MCP, golden and regression tests, human-in-the-loop vs coding loop.
- S6 · Deploy and integrate — SaaS vs console tools, MCP with Railway, containers and blast radius, worktrees, keeping the PRD aligned.
Security, confidentiality & sovereignty
- S7 · Securing the code — shift-left rules, Claude-driven code review, Project Glasswing and Mythos beyond the hype, hunting with FOIL.
- S8 · Distributing your work — Docker, publishing your own plugin to a marketplace, NPM, PyPI, Homebrew, a live Railway deploy.
- S9 · Local LLMs, with and without Claude Code — hardware, open-weight models, inference engines, hands-on on a real DGX Spark.
- Close — participants present their work; retrospective on what worked and where Claude failed.
Before you arrive
Prerequisites
- Main requirement: creativity
- Comfortable with shell, git, and either Python or TypeScript
- A concrete understanding of how an application is designed
- Offensive-security fundamentals are useful but not essential
What to bring
- A laptop with Docker, preferably Mac or Linux (Windows works with WSL)
- An active Claude Code account (Pro/Max) or API credits — or a local-model setup with
claude-code-local - Playwright installed; optionally Ollama or LM Studio
Who should attend
- Tool developers, offensive security engineers, security enthusiasts and researchers who enjoy building tools
- Anyone approaching AI-assisted spec-driven development — not for "write me a million-dollar app" vibe coding
- Level: advanced in a hacking context, medium in a development context
Taught by the maker of FOIL
The curriculum is grounded in real work, not slideware. Vito Rallo — founder of PEACH STUDIO, Cybersecurity Consult Partner at Kyndryl, and former leader of global Red Team units at Kroll and PwC — built FOIL, a SAST scanner developed largely with Claude that surfaced several zero-days, some around 22 years old. You'll learn the same patterns he used to build it.
Explore FOILBring it to your conference or team
Delivered in full at BruCON 2026 (Mechelen, Sept 21–23). Available in person, 3 days, extendable up to 5. Slides, a 50-page booklet and lab cards included.
Book this trainingQuestions about the training? info@peachstudio.be