PNC Capabilities Reference — For PNO/PNG/PNX/PNK

Purpose: Cross-engine capability awareness for optimal routing
Updated: 2026-07-05
Version: PAI 5.1.0

🎯 Quick Routing Guide

Task Type	Route To	Why
Classification/labeling	PNO (ollama/llama3.2:3b)	Free, instant, sufficient quality
JSON extraction	PNO (ollama/qwen2.5:7b)	Free, structured output
Summarization	PNO (ollama/gemma2:9b)	Free, good quality
Simple reasoning	PNC (claude/sonnet)	Best quality, fast
Complex code (multi-file)	PNK (opencode)	Multi-file architecture
Web research	PNG (antigravity)	Fast, free (Gemini)
Verification/testing	PNX (codex)	Sandboxed execution
Algorithm mode	PNC only	Full PAI orchestration

📚 PNC Skills (129 available)

Thinking & Analysis

Think — Unified debate/council router
BeCreative — Extended thinking mode
FirstPrinciples — Assumption decomposition
SystemsThinking — Causal loops, leverage points
RootCauseAnalysis — 5 Whys, Ishikawa, postmortem
ApertureOscillation — 3-pass scope shifting
IterativeDepth — Multi-angle ISC extraction
DesignThinking — 5-phase design process
Science — Hypothesis-test-analyze cycles

Research & Content

YouTube — Video search, transcript extraction
ArXiv — Paper search with AlphaXiv summaries
Research — Multi-source research orchestration
NotebookLM — Google NotebookLM integration
ResearchPipeline — YouTube + NotebookLM combined
ExtractWisdom — Content-adaptive wisdom extraction
Fabric — 240+ specialized analysis patterns
Parser — URL/file content extraction to JSON

Development

Engineer — Elite principal engineer agent
Architect — System design specialist
Designer — UX/UI design expert
Simplify — Post-change code quality review
CodeReviewer-Agency — Thorough code review
CreateCLI — TypeScript CLI generator
MCPBuilder — MCP server design specialist
BitterPillEngineering — Instruction set audit

Security

SecurityTriage — Adversarial finding verification
PAISecurityAudit — STRIDE threat modeling for PAI
InfoSecRiskAssessment — Vendor/SaaS risk assessment
WebAssessment — Web security testing
PromptInjection — Jailbreak testing
Recon — Security reconnaissance

Project & Task Management

ISA — Ideal State Artifact creation/management
PIGSD — pi-gsd project delivery framework
ProjectState — Project-scoped state management
PhaseRoadmap — Multi-phase project roadmapping
PRWorkflow — Automated pull request creation
DeployOrchestrator — Deploy→smoke test→verify→rollback
Delegation — Multi-agent parallel execution
Loop — Iterative improvement cycles
Optimize — Autonomous optimization loop

Infrastructure & Deployment

Cloudflare (ecosystem) — Workers, Pages, R2, D1, KV, Email
agents-sdk — Cloudflare Agents SDK
durable-objects — Durable Objects patterns
workers-best-practices — Production hardening
wrangler — Wrangler CLI reference
cloudflare-email-service — Email Sending + Routing
sandbox-sdk — Sandboxed code execution

Content Creation

WriteStory — Fiction writing with Storr’s method
Art — Visual content creation (Flux, GPT-Image)
Webdesign — Claude Design + frontend-design
Remotion — Programmatic video with React
AudioEditor — Whisper→Claude→ffmpeg pipeline
Excalidraw — Diagram generation
Spatial3D — 3D model from floor plans

Testing & Quality

QATester — Browser-based UI validation
UIReviewer — User story validation with Playwright
RealityChecker — Skeptical quality gate
Evals — Agent evaluation framework
UXAudit — Pre-delivery UX checklist
DesignCritique — Nielsen heuristics review

Knowledge & Memory

Knowledge — Typed graph (People/Companies/Ideas)
Telos — Life goals, projects, dependencies
Interview — Phased conversational context intake
Migrate — Content migration into PAI
ContextSearch — Search PRDs, git, sessions
WorkCommand — WORK session recall

PAI System

OllamaSkill — Local inference routing (THIS IS KEY)
AntigravitySkill — Web research dispatch
PAIUpgrade — Extract system improvements
HealthCheck — Infrastructure health probe
SettingsLab — Self-improving settings system
BitterPillEngineering — Over-prompting audit

Specialized

Ideate — Evolutionary ideation engine (9 phases)
RedTeam — 32-agent adversarial analysis
MultiEngineSynthesis — Cross-engine consolidation
WorldThreatModel — 11-horizon future stress test
TRVRM — Vulnerability lifecycle management
KeyRotation — Incident-response key rotation
ParentalControl — PAI-powered child PC management

🤖 PNC Agents (33 available)

Development

Engineer — Elite principal engineer (Fortune 10 experience)
Architect — System design specialist (distributed systems PhD-level)
Designer — UX/UI design expert (design school pedigree)
BackendDev — Backend specialist
explore — Fast codebase exploration

Quality & Testing

QATester — Browser-based functional validation (Gate 4)
RealityChecker — Skeptical quality gate (default: NEEDS WORK)
CodeReviewer-Agency — Thorough code review (blocker/suggestion/nit)
UIReviewer — Playwright user story validation

Research

ClaudeResearcher — Claude WebSearch, multi-query decomposition
GeminiResearcher — Google Gemini, 3-10 query variations
CodexResearcher — Remy, multi-model consultation (O3/GPT-5/GPT-4)
DeepResearcher — Comprehensive investigation
PerplexityResearcher — Ava, investigative analyst with Perplexity
GrokResearcher — Johannes, contrarian fact-based via xAI Grok

Specialized

Algorithm — Full PAI Algorithm runner (v5.7.11)
AgentsOrchestrator — Multi-agent pipeline manager
ProductManager — Read-only requirements analyst
MCPBuilder — MCP server design specialist
Artist — Visual content creator (Media skill workflows)
Pentester — Web security assessment (WebAssessment workflows)

Operations

IncidentResponseCommander — SEV1-SEV4 structured response
InfraOps — Docker, systemd, backup verification
DBOps — Qdrant, ICM memory, schema verification

Intelligence

Intern — 176 IQ generalist, 5 PhDs, high-agency problem solver

Browser

BrowserAgent — Parallel headless browser automation (Playwright)

🔄 Routing Decision Tree

Task received
    │
    ├─ Classification/scoring? → PNO (ollama/llama3.2:3b, free)
    ├─ JSON extraction? → PNO (ollama/qwen2.5:7b, free)
    ├─ Summarization? → PNO (ollama/gemma2:9b, free)
    │
    ├─ Web research? → PNG (antigravity, Gemini, free)
    ├─ YouTube search? → Skill('YouTube')
    ├─ Academic papers? → Skill('ArXiv')
    │
    ├─ Simple code edit? → PNC (claude/sonnet)
    ├─ Multi-file refactor? → PNK (opencode)
    ├─ Verification? → PNX (codex sandbox)
    │
    ├─ Security assessment? → Skill('SecurityTriage') or Agent('Pentester')
    ├─ Code review? → Agent('CodeReviewer-Agency') or Skill('Simplify')
    │
    ├─ Skill exists for task? → Skill('[name]')
    ├─ Agent specializes? → Agent('[type]')
    │
    └─ Complex/multi-step? → PNC Algorithm mode

💰 Cost Matrix

Engine	Model	Cost	Speed	Quality	Use For
PNO	llama3.2:3b	$0	Fast	Good	Classification
PNO	qwen2.5:7b	$0	Fast	Good	JSON
PNO	gemma2:9b	$0	Fast	Good	Summarization
PNG	Gemini	$0	Fast	Good	Web research
PNC	Sonnet	$$$	Medium	Excellent	Code, reasoning
PNK	OpenCode	$$$	Medium	Excellent	Multi-file
PNX	Codex	$$	Slow	Good	Verification

🎯 When PNO Should Escalate to PNC

Stay local (PNO) when:

Task is classification/scoring/labeling
JSON extraction with clear schema
Summarization of content
Simple categorization
Pattern matching

Escalate to PNC when:

Complex reasoning required
Code generation needed
Multi-step workflow
Skill/agent dispatch needed
Algorithm mode appropriate
ICM memory operations
Cross-engine coordination

Escalate to PNG when:

Web search required
Real-time information needed
Multi-query research
YouTube/web content lookup

Escalate to PNK when:

Multi-file architecture work
Complex refactoring
Provider-agnostic coding
TUI-driven planning needed

Escalate to PNX when:

Sandboxed verification
Adversarial testing (Cato)
Isolated execution required

📋 Capability Check Commands

List all skills:

ls ~/.claude/skills/

List all agents:

ls ~/.claude/agents/*.md

Check skill description:

grep "^description:" ~/.claude/skills/[name]/SKILL.md

Check routing policy:

cat ~/.claude/PAI/Tools/Inference.ts | grep -A 10 "task-profile"

🔧 Integration Points

For PNO to query PNC capabilities:

# Check if skill exists
[ -d ~/.claude/skills/[name] ] && echo "exists"
 
# Check if agent exists
[ -f ~/.claude/agents/[name].md ] && echo "exists"
 
# Get routing recommendation
bun ~/.claude/PAI/Tools/Inference.ts --task [profile] "[input]"

For ICM capability lookup:

# Store capability metadata
icm store --topic capabilities --key "skill:[name]" --value "{...}"
 
# Query capability
icm recall --topic capabilities --query "skill:[name]"

🎓 Learning Examples

Example 1: Classification Task

Task: “Classify this email as spam or ham”

PNO thinks:

Classification task
Check routing → ollama/llama3.2:3b
Local, free, fast, sufficient quality
Decision: Stay local, use Skill(‘OllamaSkill’, ‘—task fast_classification’)

Example 2: Multi-File Refactor

Task: “Refactor authentication across 15 files”

PNO thinks:

Multi-file code task
Check routing → complex code
PNC (single-file) vs PNK (multi-file)
15 files = multi-file architecture
Decision: Escalate to PNK via opencode run "..."

Example 3: Web Research

Task: “Find latest Claude Code features”

PNO thinks:

Web search needed
Check routing → research
PNG (Antigravity) vs Skill(‘YouTube’) vs Skill(‘ArXiv’)
General web research = PNG
Decision: Escalate to PNG via agy --dangerously-skip-permissions -p "..."

📊 Capability Coverage

Domain	Skills	Agents	Coverage
Development	15+	5	Excellent
Security	8+	2	Good
Research	10+	5	Excellent
Thinking	9+	1	Good
Infrastructure	12+	2	Good
Content	8+	2	Good
Testing	6+	3	Good
Project Mgmt	8+	2	Good

Total: 129 skills + 33 agents = 162 capabilities

🚀 Quick Reference

Most commonly needed:

Classification → Skill('OllamaSkill', '--task fast_classification')
JSON → Skill('OllamaSkill', '--task structured_json')
Web research → agy --dangerously-skip-permissions -p "..."
Code review → Agent('CodeReviewer-Agency') or Skill('Simplify')
Multi-file code → opencode run "..."
Complex task → PNC Algorithm mode

Check before routing:

Is there a skill for this? (129 options)
Is there an agent for this? (33 options)
Can PNO handle locally? (classification/JSON/summary)
Is web research needed? (PNG)
Is it multi-file? (PNK)
Else → PNC

This file should be read by PNO/PNG/PNX/PNK at startup for cross-engine capability awareness.

Location: ~/.claude/PAI/PNC_CAPABILITIES_FOR_PNO.md

Quartz 5

Explorer

PNC_CAPABILITIES_FOR_PNO

PNC Capabilities Reference — For PNO/PNG/PNX/PNK

🎯 Quick Routing Guide

📚 PNC Skills (129 available)

Thinking & Analysis

Research & Content

Development

Security

Project & Task Management

Infrastructure & Deployment

Content Creation

Testing & Quality

Knowledge & Memory

PAI System

Specialized

🤖 PNC Agents (33 available)

Development

Quality & Testing

Research

Specialized

Operations

Intelligence

Browser

🔄 Routing Decision Tree

💰 Cost Matrix

🎯 When PNO Should Escalate to PNC

📋 Capability Check Commands

🔧 Integration Points

🎓 Learning Examples

📊 Capability Coverage

🚀 Quick Reference

Graph View

Table of Contents