CategoryWeightCategory scoreContribution
Securityprompt, exec, net, exfil, eval
35%
0.0 pts
Supply chainhash, typosquat, maintainer, lockfile
20%
20.0 pts
Maintenancestaleness, pinning, CI
15%
15.0 pts
TransparencySKILL.md, perms, README
15%
15.0 pts
Communityinstalls, verify, response
15%
15.0 pts
Findings & checks · 8 flagged
Security
HIGH"Ignore previous instructions" command embedded in the skillSS-SKILL-INJECT-IGNORE-01 · Prompt injection · .claude/roles/ai-engineer/skills/04-llm-guardrails/README.md
HIGH — when it fires on hostile content the impact is full system-prompt override.
Why it matters
The text {match} is the classic direct prompt-injection phrasing. Placed in a skill body that the agent reads as trusted instructions, it tries to make the agent abandon its prior rules and follow whatever comes next — a full system-prompt override.
The exact value spotted
454self.detector = PromptInjectionDetector()
455self.attack_patterns = [
456r"ignore previous instructions",
457r"disregard all prior",
458r"new instructions:",
Occurrences
1 occurrence · at L456How to fix
Remove the override phrase, or rephrase the passage so it does not instruct the agent to discard its rules.
- Delete the
ignore/disregard/forget … previous instructionssentence. - If this is jailbreak-research or tutorial content, move the example into a clearly fenced, non-instruction block and label it as a quoted sample.
Framework references
OWASPLLM01ATLASAML.T0051Trace & refs
HIGH"Ignore previous instructions" command embedded in the skillSS-SKILL-INJECT-IGNORE-01 · Prompt injection · .claude/tech_skills/roles/ai-engineer/skills/04-llm-guardrails/README.md
HIGH — when it fires on hostile content the impact is full system-prompt override.
Why it matters
The text {match} is the classic direct prompt-injection phrasing. Placed in a skill body that the agent reads as trusted instructions, it tries to make the agent abandon its prior rules and follow whatever comes next — a full system-prompt override.
The exact value spotted
454self.detector = PromptInjectionDetector()
455self.attack_patterns = [
456r"ignore previous instructions",
457r"disregard all prior",
458r"new instructions:",
Occurrences
1 occurrence · at L456How to fix
Remove the override phrase, or rephrase the passage so it does not instruct the agent to discard its rules.
- Delete the
ignore/disregard/forget … previous instructionssentence. - If this is jailbreak-research or tutorial content, move the example into a clearly fenced, non-instruction block and label it as a quoted sample.
Framework references
OWASPLLM01ATLASAML.T0051Trace & refs
MEDIUMInstruction telling the agent not to ask for approvalSS-SKILL-INJECT-DONT-ASK-01 · Prompt injection · .claude/agents/USER-PROTOCOL.md×2
MEDIUM — it fires on intent; the real damage depends on the host agent's own approval-gating.
Why it matters
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The exact value spotted
5## Core Principle
6
7Ask when the answer changes what you'd do. Don't ask when you already know the right path.
8
9## When to Ask the User
Occurrences
2 occurrences · first at L7, also L61Show all 2 locations
Line
File
L7
.claude/agents/USER-PROTOCOL.md
L61
.claude/agents/USER-PROTOCOL.md
How to fix
Remove the approval-skipping instruction, or scope it narrowly to a specific safe, reversible action.
- Delete blanket "don't ask / no need to confirm" directives from the skill.
- If the skill is a genuine autonomous job, restrict the opt-out to a named non-destructive action rather than all actions.
Framework references
OWASPLLM01ATLASAML.T0051Trace & refs
MEDIUMInstruction telling the agent not to ask for approvalSS-SKILL-INJECT-DONT-ASK-01 · Prompt injection · .claude/agents/specialists/brainstorm-architect-agent.md×2
MEDIUM — it fires on intent; the real damage depends on the host agent's own approval-gating.
Why it matters
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The exact value spotted
38**Step 2: Ask Targeted Questions (Maximum 5)**
39
40Ask ONLY questions where the answer changes your recommendation. Never ask for information y
… (41 chars elided on L40)
41
42```
Occurrences
2 occurrences · first at L40, also L45Show all 2 locations
Line
File
L40
.claude/agents/specialists/brainstorm-architect-agent.md
L45
.claude/agents/specialists/brainstorm-architect-agent.md
How to fix
Remove the approval-skipping instruction, or scope it narrowly to a specific safe, reversible action.
- Delete blanket "don't ask / no need to confirm" directives from the skill.
- If the skill is a genuine autonomous job, restrict the opt-out to a named non-destructive action rather than all actions.
Framework references
OWASPLLM01ATLASAML.T0051Trace & refs
MEDIUMInstruction telling the agent not to ask for approvalSS-SKILL-INJECT-DONT-ASK-01 · Prompt injection · .claude/skill-docs/brainstorm-architect.md
MEDIUM — it fires on intent; the real damage depends on the host agent's own approval-gating.
Why it matters
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The exact value spotted
45- "Is this replacing an existing system?" (changes migration strategy)
46
47BAD questions (don't ask these):
48- "What language do you want?" (check the project — it's already decided)
49- "Should I use TypeScript?" (check existing config files)
Occurrences
1 occurrence · at L47How to fix
Remove the approval-skipping instruction, or scope it narrowly to a specific safe, reversible action.
- Delete blanket "don't ask / no need to confirm" directives from the skill.
- If the skill is a genuine autonomous job, restrict the opt-out to a named non-destructive action rather than all actions.
Framework references
OWASPLLM01ATLASAML.T0051Trace & refs
MEDIUMInstruction telling the agent not to ask for approvalSS-SKILL-INJECT-DONT-ASK-01 · Prompt injection · CLAUDE.md
MEDIUM — it fires on intent; the real damage depends on the host agent's own approval-gating.
Why it matters
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
The exact value spotted
112- **Report AFTER milestones** with progress summaries at key checkpoints
113- **Report AFTER completion** with final summary and next steps
114- **Don't ask** about obvious implementation details, style choices, or decisions within age
… (12 chars elided on L114)
115
116Reference: `.claude/agents/USER-PROTOCOL.md` for the full protocol.
Occurrences
1 occurrence · at L114How to fix
Remove the approval-skipping instruction, or scope it narrowly to a specific safe, reversible action.
- Delete blanket "don't ask / no need to confirm" directives from the skill.
- If the skill is a genuine autonomous job, restrict the opt-out to a named non-destructive action rather than all actions.
Framework references
OWASPLLM01ATLASAML.T0051Trace & refs
Supply chain
✓—All supply chain checks passedNo findings in this category for the latest scan.pass
Maintenance
✓—All maintenance checks passedNo findings in this category for the latest scan.pass
Transparency
✓—All transparency checks passedNo findings in this category for the latest scan.pass
Community
✓—All community checks passedNo findings in this category for the latest scan.pass
Every scanned point with the score it earned and what moved between them.
ae42a04→ae42a04100 → 65-35
Score fell 35 points between these scans.
Computing diff…
The primary manifest — the file an agent reads to learn what this artifact does.
Tech Hub Skills
200+ production-ready AI agent skills for Claude Code and GitHub Copilot.
Features a hierarchical multi-agent system with a Brainstorm → Plan → Implement workflow, 95% token efficiency, advanced domain expertise, and comprehensive safety guardrails for damage control.
Quick Install
You can install this system directly from GitHub. No manual cloning is required for the primary installation methods below.
1. For Claude Code (Agentic Orchestration)
Install the plugin directly inside Claude Code using one of these two methods:
Option A: Fast CLI (Recommended) Run these commands inside your Claude Code session:
/plugin marketplace add yakoub-ai/Tech-Skills
/plugin install tech-skills@yakoub-aiOption B: Interactive Marketplace UI
- Open Claude Code and type
/plugin - Go to the Marketplaces tab → + Add Marketplace
- Enter the repo URL:
https://github.com/yakoub-ai/Tech-Skills - Go to the Discover tab → find
tech-skills→ select Install for you - Restart Claude Code
2. For GitHub Copilot (VSCode)
Add expert agent skills to your project in one command. You do not need to clone the repository first:
npx yakoub-ai/Tech-Skills install --copilot_This creates .github/copilot-instructions.md that provides expert context to Copilot._
3. Advanced: Manual Setup
If you prefer to include the agent system as part of your source code, you can clone it manually:
git clone https://github.com/yakoub-ai/Tech-Skills.git
cp -r Tech-Skills/.claude/ /your/project/
cp Tech-Skills/CLAUDE.md /your/project/v3.0: Agent Orchestration Framework
What's new in v3.0:
- CLAUDE.md bootstrap — System activates automatically when Claude Code opens the project
- Context-first protocol — Agents explore and understand your project before making changes
- Real agent execution — Agents use Claude Code's native Agent tool for true subagent spawning
- Parallel execution — Independent tasks run simultaneously for faster completion
- Quality gate enforcement — Mandatory verification before any task is marked complete
- User communication protocol — Agents ask the right questions at the right time
- Feedback loops — Results are validated and iterated on if quality gates fail
- marketplace.json — Easy discovery and installation metadata
Multi-Agent Hierarchy
User Request
|
Orchestrator (analyzes, plans, coordinates)
|
5 Lead Agents (AI/ML, Platform, Security, Data, Product)
|
25+ Specialist Agents (each with 5-13 skills, loaded on-demand)
|
200+ Skills (expert guidance loaded from skill-docs/)Workflow: Brainstorm → Plan → Implement
- Brainstorm: Explores the project, understands requirements, identifies constraints and risks, asks clarifying questions
- Plan: Scans skill registries, selects minimum specialists (3-7), defines milestones, presents plan for approval
- Implement: Spawns agents with full context, executes in parallel where possible, validates quality gates, synthesizes results
Usage
Slash Commands
/orchestrator "Build a customer churn prediction model" # Full workflow
/ai "Create a RAG chatbot" # AI/ML tasks
/platform "Deploy to Kubernetes with CI/CD" # DevOps/cloud
/security "Scan for PII and vulnerabilities" # Security
/data "Build ETL pipeline with quality checks" # Data engineering
/product "Create user registration with tests" # Full-stack devNatural Language
Just describe what you need — the orchestrator automatically activates, analyzes keywords, and routes to the right agents:
"Build a REST API with authentication, rate limiting, and monitoring"
→ Orchestrator → Product Lead (backend) + Security Lead (auth) + Platform Lead (monitoring)GitHub Copilot
Reference roles in comments to steer Copilot:
# Using AI Engineer approach for RAG pipeline
def build_rag_system():
pass
# Apply Security Architect best practices
def handle_user_upload(file_data):
passAvailable Roles
| Lead | Domain | Specialists | Skills |
|---|---|---|---|
| AI/ML Lead | AI, ML, Data Science | AI Engineer, ML Engineer, Data Scientist, MLOps | 39 |
| Platform Lead | Infrastructure, DevOps | DevOps, SRE, Platform Eng, Network, Docker, AWS/Azure/GCP, FinOps | 83 |
| Security Lead | Security, Compliance | Security Architect, Compliance Officer, Security Hardener | 23 |
| Data Lead | Data Engineering | Data Engineer, Data Governance, Database Admin | 26 |
| Product Lead | Product Development | Product Designer, Frontend/Backend Dev, QA, Tech Writer | 38 |
Key Features
- 200+ Skills across LLMs, RAG, MLOps, DevSecOps, Lakehouse, Cloud (AWS/Azure/GCP), and more
- 95% Token Efficiency via lazy-loading skill registries — only loads what's needed
- Context-First — Agents analyze your project structure, tech stack, and conventions before acting
- Security First — Built-in PII detection, security hardening, and compliance automation
- Quality Gates — Mandatory verification including tests, linting, and security checks
- Safety Guardrails — File deletion protection, database safety, credential protection, audit logging
- Parallel Execution — Independent agents work simultaneously for faster task completion
- Works on Any Project — Agents adapt to your existing conventions, no configuration needed
Project Structure
.claude/
agents/ # Agent definitions (orchestrator, leads, specialists)
orchestrator-agent.md # Master coordinator
ai-ml-lead.md # AI/ML domain lead
platform-lead.md # Platform/DevOps lead
security-lead.md # Security/compliance lead
data-lead.md # Data engineering lead
product-lead.md # Product development lead
specialists/ # 25+ specialist agents
EXECUTION.md # How agents coordinate
SKILL-REGISTRY.md # Skill keyword index
ROLE-REGISTRY.md # Role summary index
QUALITY-GATES.md # Verification checklist
CONTEXT-PROTOCOL.md # Project analysis protocol
USER-PROTOCOL.md # User communication rules
commands/ # Slash command definitions
skill-docs/ # Expert guidance (loaded on-demand)
roles/ # Detailed skill implementations
hooks/ # Safety guardrails
settings.json # System configuration
CLAUDE.md # Auto-loaded bootstrap
marketplace.json # Installation metadataDocumentation
- Architecture: AGENTS.md
- Safety Guardrails: SAFETY-GUARDRAILS.md
- GitHub Copilot: GITHUB_COPILOT.md
- Changelog: CHANGELOG.md
License
MIT