Deep Skills Suite

A collection of composable, model-agnostic, domain-agnostic AI agent skills for multi-agent analysis, verification, research, and council-of-councils review.

Overview

Skills are conversation-driven — they extract context from what you've been discussing and act on it. No configuration or keywords needed.

Architecture principles:

• Inline execution — context-sensitive skills run in the same agent, inheriting full conversation history
• Isolated execution — stateless data-retrieval skills (brave-search, perplexity, deepwiki) run as isolated sub-agents
• Progressive enhancement — each skill works standalone; multi-agent debate, model diversity, runtime diversity, and cross-council synthesis are additive layers
• Non-blocking bridges — unavailable CLI tools produce SKIPPED, never halt execution

Installing Skills

Via skillsmp.com

Browse and install skills from the registry at skillsmp.com.

Via skills.sh

# Install individual skills
npx skills add mikeng-io/agent-skills --skill deep-review
npx skills add mikeng-io/agent-skills --skill deep-audit
npx skills add mikeng-io/agent-skills --skill deep-verify
npx skills add mikeng-io/agent-skills --skill deep-research
npx skills add mikeng-io/agent-skills --skill deep-explorer
npx skills add mikeng-io/agent-skills --skill deep-council

# Install the full suite at once
npx skills add mikeng-io/agent-skills --all

Manual Installation

# Clone the repo
git clone https://github.com/mikeng-io/agent-skills
cd agent-skills

# Required foundation (deep-* skills depend on these)
ln -s $(pwd)/skills/domain-registry ~/.claude/skills/domain-registry
ln -s $(pwd)/skills/context ~/.claude/skills/context
ln -s $(pwd)/skills/preflight ~/.claude/skills/preflight
ln -s $(pwd)/skills/parallel-workflow ~/.claude/skills/parallel-workflow
ln -s $(pwd)/skills/debate-protocol ~/.claude/skills/debate-protocol
ln -s $(pwd)/skills/council-taxonomy ~/.claude/skills/council-taxonomy

# Deep skills
ln -s $(pwd)/skills/deep-explorer ~/.claude/skills/deep-explorer
ln -s $(pwd)/skills/deep-review ~/.claude/skills/deep-review
ln -s $(pwd)/skills/deep-audit ~/.claude/skills/deep-audit
ln -s $(pwd)/skills/deep-verify ~/.claude/skills/deep-verify
ln -s $(pwd)/skills/deep-research ~/.claude/skills/deep-research
ln -s $(pwd)/skills/deep-council ~/.claude/skills/deep-council

# Optional — council-of-councils bridges
ln -s $(pwd)/skills/bridge-commons ~/.claude/skills/bridge-commons
ln -s $(pwd)/skills/bridge-claude ~/.claude/skills/bridge-claude
ln -s $(pwd)/skills/bridge-gemini ~/.claude/skills/bridge-gemini
ln -s $(pwd)/skills/bridge-codex ~/.claude/skills/bridge-codex
ln -s $(pwd)/skills/bridge-opencode ~/.claude/skills/bridge-opencode

Quick Reference

Example workflows:

# Understand a new codebase
/deep-explorer

# Pre-deployment
/deep-audit      # Standards compliance
/deep-verify     # Risk verification

# Research → Implement → Review
/deep-research   # Background research
# ... implement ...
/deep-review     # Improvement suggestions
/deep-council    # Cross-runtime/cross-council confidence boost

Skills

Foundation

context

Classifies the current conversation artifact (code, financial, marketing, etc.), selects domains from the domain registry, and recommends routing (parallel-workflow, debate-protocol, or deep-council). Invokable standalone to inspect how the suite reads your current conversation.

preflight

Lightweight scope clarifier. Invoked by deep-* skills when conversation context is too sparse to determine what to analyze. Asks 1–3 targeted questions one at a time (one per message, multiple-choice preferred), then returns a structured scope_clarification block for the calling skill. Skipped automatically when scope is already clear. Adapted from the superpowers brainstorming skill — same principle of asking before acting, scoped to analysis intent rather than feature design.

debate-protocol

5-phase structured adversarial debate. Domain experts analyze independently, a Devil's Advocate challenges every CRITICAL/HIGH finding, an Integration Checker surfaces cross-component gaps, findings are confirmed/withdrawn/disputed/merged. Also supports brainstorm/design proposal lifecycles, packetized council exchange, and nested-council artifacts. Used by deep-council and deep-verify.

council-taxonomy

Shared vocabulary for Agent Council, Model Council, Runtime Council, and Deep Council. Use before designing or invoking council workflows so role, model, runtime/toolchain, evidence-channel, and debate-layer diversity are not conflated.

parallel-workflow

DAG-based parallel dispatch. Spawns independent sub-agents, collects results, synthesizes. Used by all deep-* skills as their default execution path.

domain-registry

Reference library of domain definitions (technical, business, creative). Skills read this to select the right expert roles for any artifact type. Not invokable — read via the Read tool.

Deep Skills

deep-explorer

Explores and maps a codebase or artifact. Git-aware: full exploration on first run, delta tracking on subsequent runs.

Output: .outputs/exploration/

deep-review

Multi-agent quality improvement review. Spawns domain-appropriate reviewers in parallel. Returns prioritized suggestions — not pass/fail, only improvement-focused.

Output: .outputs/review/

deep-audit

Standards and compliance audit with formal verdicts. Auditor selection driven by domain-registry — a marketing artifact gets brand/legal auditors, a healthcare app gets HIPAA auditors.

Verdict: PASS / CONCERNS / FAIL
Output: .outputs/audit/

deep-verify

Risk-focused verification with Devil's Advocate as a mandatory challenger. DAG execution: domain experts → (DA + Integration Checker in parallel) → Third-Party Reviewer. Optional council-of-councils second pass via deep-council.

Verdict: PASS / CONCERNS / FAIL
Output: .outputs/verification/

deep-research

Multi-domain research with parallel domain researchers. Automatic tool discovery — uses Brave Search, Perplexity, web reader, browser automation, DeepWiki, and documentation query tools as available. Cross-domain synthesis identifies insights that emerge from intersections.

Output: .outputs/research/

Council-of-Councils

deep-council

Dispatches discovery-first packets to all available council-capable runtimes in parallel. Each bridge runs the richest local council it supports, then Deep Council synthesizes proposals/findings through cross-council debate. Two-layer debate architecture:

• Layer 2 (intra-bridge): Each bridge extracts maximum value from its own runtime/toolchain/model setup through local Agent Council, Model Council, or debate-compatible review
• Layer 1 (cross-bridge): Debate Coordinator challenges findings/proposals across all bridges — multi-source confirmation gets elevated, conflicts and shared-bias risks get flagged

Bridges:

Output: .outputs/council/

Optional Data Sources

These run as isolated sub-agents (context: fork) — stateless lookups, no conversation context needed.

deepwiki

Queries a Devin DeepWiki index for a GitHub repository. Returns AI-grounded answers about codebase architecture, component relationships, and design decisions. Requires Devin API key.

claude mcp add -s user -t http devin https://mcp.devin.ai/mcp -H "Authorization: Bearer <API_KEY>"

brave-search

Brave Search MCP wrapper. Web, news, and local search with an independent index. Requires Brave Search API key.

claude mcp add -s user brave-search npx @brave/search-mcp
# BRAVE_API_KEY=your_key

perplexity

Perplexity MCP wrapper. Returns AI-synthesized answers with inline citations — best for consensus questions and cross-validation. Requires Perplexity API key.

claude mcp add -s user perplexity npx @perplexity-ai/mcp-server
# PERPLEXITY_API_KEY=your_key

Repository Structure

agent-skills/
├── README.md
├── .bridge-settings.json          # Suite config: bridge toggles, opencode models, ttl
│
└── skills/
    ├── TOPOLOGY.md                # Full architecture map, routing decisions, dependency graph
    │
    ├── domain-registry/           # Reference library of domain definitions
    │   ├── README.md
    │   └── domains/
    │       ├── technical.md
    │       ├── business.md
    │       └── creative.md
    │
    ├── bridge-commons/            # Shared contract for all bridge adapters
    │   └── SKILL.md
    │
    ├── debate-protocol/           # 5-phase adversarial debate + brainstorm/proposal protocol
    │   ├── SKILL.md
    │   ├── experts/
    │   │   ├── devils-advocate.md
    │   │   ├── integration-checker.md
    │   │   ├── test-architect.md
    │   │   └── third-party.md
    │   └── schemas/
    │
    ├── context/                   # Artifact classifier and smart router
    │   └── SKILL.md
    ├── council-taxonomy/          # Agent/Model/Runtime/Deep Council reference
    │   └── SKILL.md
    │
    ├── parallel-workflow/         # DAG-based parallel agent dispatcher
    │   └── SKILL.md
    │
    ├── bridge-claude/             # Claude Code sub-agent bridge
    │   └── SKILL.md
    ├── bridge-gemini/             # Gemini CLI bridge
    │   └── SKILL.md
    ├── bridge-codex/              # Codex CLI/MCP bridge
    │   └── SKILL.md
    ├── bridge-opencode/           # OpenCode multi-model bridge
    │   └── SKILL.md
    │
    ├── deep-council/              # Council-of-councils orchestrator
    │   ├── SKILL.md
    │   ├── README.md
    │   └── schemas/
    │
    ├── deep-explorer/             # Git-aware codebase exploration
    │   ├── SKILL.md
    │   └── README.md
    │
    ├── deep-review/               # Quality improvement review
    │   ├── SKILL.md
    │   ├── README.md
    │   └── schemas/
    │
    ├── deep-audit/                # Standards compliance audit
    │   ├── SKILL.md
    │   ├── README.md
    │   └── schemas/
    │
    ├── deep-verify/               # Risk verification with DA
    │   ├── SKILL.md
    │   ├── README.md
    │   ├── schemas/
    │   └── validators/
    │
    ├── deep-research/             # Multi-domain research
    │   ├── SKILL.md
    │   ├── README.md
    │   ├── schemas/
    │   └── validators/
    │
    ├── deepwiki/                  # Devin DeepWiki data source
    │   └── SKILL.md
    ├── brave-search/              # Brave Search MCP data source
    │   └── SKILL.md
    └── perplexity/                # Perplexity MCP data source
        └── SKILL.md

Skill Execution Model

Skill type          context field    Conversation inherited?
─────────────────   ─────────────    ──────────────────────
Deep skills         (none — inline)  Yes — sees full history
Foundation          (none — inline)  Yes — sees full history
Data sources        context: fork    No — stateless lookup
Bridge adapters     context: reference  N/A — read via Read tool, never invoked

Configuration

.bridge-settings.json

Suite-level configuration at the repo root:

{
  "bridges": {
    "claude":   { "enabled": true },
    "gemini":   { "enabled": true },
    "codex":    { "enabled": true, "model": null },
    "opencode": { "enabled": true, "model": null, "models": [] }
  },
  "reasoning_level": "medium",
  "ttl_hours": 24
}

opencode.models — set to 2+ model strings to enable multi-model dispatch within bridge-opencode:

"models": ["anthropic/claude-sonnet-4-20250514", "openai/gpt-4o", "google/gemini-2.0-flash"]

Design Principles

1. Non-blocking — a missing CLI never halts execution; it produces SKIPPED
2. Conversation-driven — no triggers, keywords, or configuration required
3. No hardcoded models — capability levels (highest, high, standard) only
4. Progressive enhancement — each skill works at minimum capability; debate, local councils, model diversity, and runtime/toolchain diversity are additive
5. Domain-agnostic — domain-registry drives expert selection; the same skills work for code, finance, marketing, design, legal
6. Composable — any skill can be used standalone or as part of a larger orchestration

Contributing

1. Fork the repository
2. Create your skill under skills/your-skill/
3. Follow the skill structure: SKILL.md (required), README.md (required), schemas/ (optional)
4. Use capability levels, not model names
5. No context: fork unless the skill is stateless and needs no conversation history
6. Submit a pull request

Skill checklist:

• SKILL.md with valid frontmatter (name, description, location, allowed-tools)
• README.md with purpose, features, quick start
• Model-agnostic (capability levels only)
• Domain-agnostic (domain-registry or conversation-driven selection)
• Output to .outputs/{purpose}/ with YAML frontmatter and JSON companion
• context: fork only if stateless (no conversation context needed)