customize
A Claude Code skill that ports the vuln-pipeline to your stack: a new language, vulnerability class, detection mechanism, or target shape. It reads the pipeline source, interviews you about your target, presents a migration plan, and executes the edits after you approve. For when to reach for it and what porting involves, see docs/customizing.md.
Requirements
- Claude Code CLI installed and authenticated
- This repo checked out — the skill edits the pipeline in place (the
domain-specific parts of
harness/: stage prompts, trace parser, artifact schemas) - Docker + the sandbox set up (
scripts/setup_sandbox.sh) for the validation run at the end
Usage
Inside Claude Code:
> /customizeThe skill runs in five steps: it reads the pipeline source, interviews you (operating context and goal first, then technical follow-ups: detection signal, PoC shape, grading criteria), presents a file-by-file plan, waits for your explicit approval, then executes and validates.
Setup variants — including running the port outside Claude Code — are covered in docs/customizing.md.
Output
No run directory — the output is the port itself:
- A migration plan you approve before any edit (what changes, what stays, open questions)
- Edits to the domain-specific files: the stage prompts under
harness/prompts/, the trace parser, artifact schemas, and the target Dockerfile template - A canary target under
targets/<domain>-canary/with planted bugs of the new class, plus a validation run confirming the ported pipeline finds and grades them
Questions
Reach out to your Anthropic contact.