Red-team expertise as Claude system prompts

What it does

claude-red is a curated collection of SKILL.md files for Anthropic’s Claude Skills system. Each file primes the model with expert-level methodology for a specific attack surface — SQL injection, WPA3 side-channels, EDR evasion, ADCS abuse, and so on. Drop them into Claude’s skills directory and the model loads relevant expertise on demand, triggered by what you’re discussing, without burning context on skills you aren’t using.

The interesting bit

The project treats prompt engineering as a knowledge-curation problem rather than a one-off craft exercise. Someone has already done the work of encoding 13 categories of offensive tradecraft — from Zigbee/Thread/Matter to exploit development — into structured, reusable system context. It’s essentially a field manual that Claude can page through automatically.

Key highlights

• 58 skills across 13 categories, with wireless (13 skills) and web application (16 skills) the most mature
• Covers niche terrain: LoRaWAN/sub-GHz, BLE pairing downgrade, KRACK/FragAttacks, RAG poisoning
• Sparse-checkout support lets you install single categories instead of the full repo
• Active Directory, cloud, mobile, and IoT categories are currently thin (1 skill each) but have explicit roadmaps for expansion into 10–16 focused sub-skills
• Includes utility skills for fast-checking and professional reporting

Caveats

• Several categories are explicitly marked “expanding” with only overview-level skills currently available
• The README is upfront that this is for authorized engagements, but the safety boundary is left to the operator and platform policy, not technical enforcement

Verdict

Worth bookmarking if you’re a red teamer or bug bounty hunter already using Claude Code or the Claude Skills system. If you’re not in that workflow, this is just a well-organized set of markdown files you can’t easily use elsewhere.