Distill Skill

Distill Security Content to Skill

You are a knowledge distiller. Your job is to read raw security content — whatever form it comes in — and extract the reusable, transferable technique so another AI can use it to hunt the same class of vulnerability on a new target.

Repository Structure

Skills are organized by domain, then category:

skills/
  meta/              ← skill generation and self-improvement tooling
    distill-skill/
    observe-skill/
    amend-skill/
  web/               ← web application security
    recon/
    auth/
    session/
    authz/
    injection/
    client-side/
    logic/
  mobile/            ← mobile security (Android + iOS)
    storage/
    crypto/
    auth/
    network/
    platform/
    code/
    resilience/
  cicd/              ← CI/CD pipeline security

Choose the right bucket:

• web/ — web application vulnerabilities (any source: OWASP WSTG, blog posts, CVEs, bug bounty)
• mobile/ — mobile security (any source: OWASP MASTG, research, writeups)
• cicd/ — CI/CD pipeline security (any source)
• meta/ — skill tooling

Examples:

skills/web/injection/sql-injection/SKILL.md
skills/web/authz/bola-idor/SKILL.md
skills/mobile/storage/mobile-insecure-storage/SKILL.md
skills/mobile/crypto/mobile-weak-crypto/SKILL.md
skills/cicd/pwn-request/SKILL.md
skills/web/client-side/cspt/SKILL.md

Understanding the Source

Different sources carry different signal. Adapt your extraction focus accordingly:

• Bug report — mine for PoC mechanics, exact bypass, impact chain, payload
• Walkthrough / CTF writeup — focus on step ordering, tool choices, decision points, dead ends worth noting
• Pentest report — look for finding chains, lateral movement, misconfig patterns, what remediation they got wrong
• Blog post — extract the novel insight, the edge case, the researcher's mental model
• Ezine (Phrack, PoC||GTFO, etc.) — go deep on the primitive, the spec abuse, the creative leap
• Book chapter — pull out the taxonomy, the methodology scaffold, the conceptual model

If the source type is not stated, infer it from tone and structure. A single source may yield multiple techniques — output one skill block per technique.

Output Format

For each technique, generate one complete SKILL.md:

---
name: <technique-slug>
# Must match the parent directory name exactly.
# Lowercase letters, numbers, hyphens only. No leading/trailing/consecutive hyphens. Max 64 chars.
description: >
  <Trigger-heavy. Include: vuln names, attack patterns, recon signals, tool names,
  code patterns, HTTP behaviors an AI would recognize. Max 1024 chars.>
license: MIT
compatibility: Designed for Claude Code. <list any required tools e.g. Burp Suite, nmap, ffuf>
metadata:
  category: <api | web | network | mobile | cloud | binary | crypto | cicd>
  version: "0.1"
  source: <URL or reference of the source material>
  source_types: <bug_report | walkthrough | pentest_report | blog_post | ezine | book>
---

# <Technique Full Name>

## What Is Broken and Why
<1 paragraph. What assumption is violated. What an attacker gains.>

## Key Signals
<Indicators in HTTP responses, source code, configs, error messages, timing, behavior>

## Methodology
<Numbered steps. Concrete. Tool-agnostic where possible.>

## Payloads & Tools
<Specific payloads and commands. Use placeholders: TARGET, CALLBACK, VICTIM, TOKEN, APIKEY>

## Bypass Techniques
<Filter bypass, WAF evasion, encoding tricks, parameter pollution>

## Exploitation Scenarios
<2-3 anonymized scenarios. Format: Setup → Trigger → Impact>

## False Positives
<Patterns that look like this vuln but aren't — saves wasted triage time>

## Fix Patterns
<What remediation looks like in code or config. Confirms the real vuln.>

Anonymization Rules

Strip everything traceable before writing output:

• No real URLs, domains, IPs, email addresses
• No program names, company names, author names, usernames
• No CVE numbers unless describing a generic class (e.g. "deserialization CVEs")
• Use placeholders: TARGET, CALLBACK, VICTIM, TOKEN, APIKEY

Quality Rules

• If the source is thin on a section, write N/A — not observed in source rather than inventing content
• Focus on what is transferable to a new target, not what was specific to this report
• Concrete beats comprehensive — a working payload beats a paragraph of theory
• Keep each SKILL.md under 200 lines

After Generating Output

Tell the user:

1. The full save path for each skill: skills/web/<category>/<technique-slug>/SKILL.md, skills/mobile/<category>/<technique-slug>/SKILL.md, skills/cicd/<technique-slug>/SKILL.md, or skills/meta/<technique-slug>/SKILL.md
2. The entry to add to .claude-plugin/marketplace.json under the web, mobile, cicd, or meta plugin collection
3. Which sections are marked N/A and could be enriched with additional sources