AI Knowledge Skill

Production-grade engineering standards for AI-assisted software development

This skill provides access to a comprehensive engineering knowledge base designed to prevent "silent failures" in AI-generated code. It contains battle-tested architectural patterns, security guidelines, database concurrency best practices, and performance optimization strategies.

Critical Problem Addressed: 40-45% of AI-generated code contains silent failures that traditional testing cannot catch.

🧠 COGNITIVE PROTOCOL (The 5-Step Deep Thinking)

Execute BEFORE generating any code or solution

Before implementing features, you MUST internally process these 5 strategic questions. This transforms you from a "code writer" to a "software architect."

Step 1: Intent Analysis ("What & Why")

• •Question: What exactly is the user asking for, and what is the business value?
• •Action: Identify the core requirement vs. the underlying need
• •Example: "User needs login" → "Business needs secure identity management"

Step 2: Future-Proofing ("What's Next")

• •Question: What features will logically follow this implementation?
• •Action: Design extensible solutions, avoiding "dead-end" architectures
• •Example: Login → 2FA → RBAC → Profile Management → Session Management

Step 3: Standardization ("How to do it Right")

• •Question: Which industry standards and Knowledge Base files apply here?
• •Action: Map the request to relevant knowledge files
• •Check: Anti-patterns in Contextual_Architecture_Decision_Matrix

Step 4: Systemic Impact ("Interactions")

• •Question: How does this interact with other systems (DB, Cache, Frontend, queues)?
• •Action: Consider side effects like race conditions, thundering herd, cascading failures
• •Reference: Architectural_Resilience for failure modes

Step 5: Validation Strategy ("How to Verify")

• •Question: How will we prove this works beyond happy-path testing?
• •Action: Plan for observability, logging, edge cases, and property-based testing
• •Consider: Security vulnerabilities, performance under load, error scenarios

⚠️ OVERENGINEERING PREVENTION (Right-Sizing Check)

CRITICAL WARNING: Do NOT apply every pattern from the Knowledge Base indiscriminately. A simple Todo app does NOT need microservices, distributed caching, or complex security layers.

Before Implementing, Ask: "How Much Is Actually Needed?"

Reference: Contextual_Architecture_Decision_Matrix_Knowledge_Base.md

The "Complexity vs. Need" Assessment

For every feature or pattern you consider applying, evaluate:

Red Flags for Overengineering

❌ STOP if you find yourself:

• •Adding Kubernetes for a prototype
• •Implementing CQRS for a CRUD app
• •Using microservices for < 1000 users
• •Adding distributed caching for < 100 RPS
• •Implementing event sourcing for simple state
• •Creating abstractions "just in case"

The "YAGNI" Principle (You Ain't Gonna Need It)

Rule: If the feature isn't needed RIGHT NOW, don't build it.

Examples:

• •❌ "I'll add Redis caching now for when we scale to 1M users"
• •✅ "Simple in-memory cache is enough for current 100 users"
• •❌ "I'll implement full RBAC with 10 permission levels"
• •✅ "Basic admin/user roles are sufficient for MVP"

Context-Aware Decision Matrix

Step 1: Assess Project Context

• •Current user count?
• •Expected growth timeline?
• •Criticality of downtime?
• •Team size and expertise?
• •Budget constraints?

Step 2: Match Solution to Context

Startup MVP (0-100 users)     → Monolith + SQLite
Growing Product (1K-10K)      → Monolith + PostgreSQL
Scale-up (10K-100K)          → Modular Monolith + Redis
Enterprise (100K+)           → Microservices + Full Stack

Step 3: Document the Trade-off When choosing a simpler solution, document:

• •Why simpler approach was chosen
• •When to upgrade (triggers)
• •Technical debt created
• •Migration path for future

Practical Examples

Example 1: Todo App

❌ OVER-ENGINEERED:
   - Kubernetes deployment
   - Microservices (AuthService, TodoService, UserService)
   - Redis caching
   - CQRS pattern
   - Event sourcing

✅ RIGHT-SIZED:
   - Single monolithic app
   - SQLite database
   - In-memory session storage
   - Simple REST API

Example 2: E-commerce MVP

❌ OVER-ENGINEERED:
   - Event-driven architecture
   - Distributed transaction saga pattern
   - Multi-region deployment
   - ML-powered recommendation engine

✅ RIGHT-SIZED:
   - Monolithic architecture
   - PostgreSQL database
   - Basic caching
   - Simple collaborative filtering

Decision Checklist

Before adding ANY architectural pattern, verify:

• • Is this needed NOW? (Not "might need in future")
• • Does complexity match project stage? (MVP vs. Production)
• • Is team capable of maintaining this? (Learning curve vs. benefit)
• • Is the cost justified? (Time, money, complexity)
• • Can we start simple and evolve? (Evolution vs. Big Bang)

When in Doubt, Start Simple

Golden Rule: You can always add complexity later. You cannot easily remove it.

Prefer:

• •Working simple code over perfect complex code
• •Concrete requirements over imagined future needs
• •Team understanding over architectural purity
• •Ship fast over build perfect

🤖 SUBAGENT KNOWLEDGE RETRIEVAL PROTOCOL

CRITICAL: DO NOT load full KB files into context. Use subagents instead.

The Context Window Problem

Knowledge Base files are LARGE (thousands of tokens each). Loading them directly into your context will:

• •❌ Bloat context window - Leaves less room for actual coding
• •❌ Increase token costs - Expensive and inefficient
• •❌ Slow down responses - More tokens = slower processing
• •❌ Cause truncation - Important details may be lost

Solution: Knowledge Retrieval Subagent

Instead of reading KB files directly, spawn a subagent to do it for you.

When to Use Subagents

ALWAYS use subagent when:

• •Reading any file from knowledgeAI/ directory
• •Need specific section from large KB file
• •Context window is already >50% full
• •Multiple KB files need to be referenced

NEVER read directly when:

• •File is >1000 tokens
• •Multiple files needed
• •Context is constrained

Subagent Task Template

@general
Task: Read and summarize relevant sections from Knowledge Base

Files to check:
- knowledgeAI/Client_Side_Security_Architecture_Knowledge_Base.md
- knowledgeAI/Comprehensive_Architectural_Framework_Knowledge_Base.md

Keywords from user request: "JWT", "token storage", "frontend"

Instructions:
1. Search for relevant sections matching keywords
2. Extract ONLY the specific patterns/anti-patterns
3. Return concise summary (max 500 tokens):
   - Key patterns to apply
   - Critical anti-patterns to avoid
   - Code examples if relevant
   - Specific file sections for reference

DO NOT return full file content. Only relevant extracts.

Example Workflow

❌ WRONG (Context Bloat):

User: "How should I store JWT tokens?"
AI: [Reads entire Client_Side_Security_Architecture_Knowledge_Base.md - 5000 tokens]
    [Context now 70% full]
    [Less room for actual solution]

✅ RIGHT (Subagent Pattern):

User: "How should I store JWT tokens?"
AI: [Spawns subagent with specific task]
    Subagent: [Reads KB, extracts JWT section only - 300 tokens]
    Subagent returns: "Use HttpOnly cookies. Avoid localStorage due to XSS."
    AI: [Uses 300 token summary to generate solution]
    [Context stays clean, efficient, fast]

Subagent Output Format

Subagent should return:

## Relevant Patterns
- [Pattern name]: [One-line description]
- [Pattern name]: [One-line description]

## Anti-Patterns to Avoid
- [Anti-pattern]: [Why it's bad]

## Quick Reference
- File: [KB filename]
- Section: [Specific section/heading]
- Key Takeaway: [One-line actionable advice]

Token Budget Rules

• •Subagent task description: Max 200 tokens
• •Subagent return summary: Max 500 tokens
• •Your context after subagent: Should be <60% full
• •Multiple KB files: Spawn multiple subagents in parallel, not sequentially

Context Management Checklist

Before reading ANY file:

• • Check current context usage
• • If >50% full → Use subagent
• • If file >1000 tokens → Use subagent
• • If multiple files needed → Use subagent
• • Keep summaries, not full text

Benefits of Subagent Pattern

✅ Efficient context usage - Only relevant info in main context ✅ Faster responses - Less token processing ✅ Cost effective - Minimal token waste ✅ Parallel processing - Multiple subagents can work simultaneously ✅ Clean architecture - Separation of concerns

💾 MEMORY PROTOCOL (Architecture Decision Records)

Execute AFTER implementing any significant feature

After implementing features, you MUST maintain persistent memory of architectural decisions. CRITICAL: Keep entries CONCISE to avoid token bloat.

⚠️ TOKEN OPTIMIZATION RULE

Memory entries must be SHORT and SCANNABLE:

• •Max 2-3 lines per field
• •Use bullet points, not paragraphs
• •Reference KB files by name only (don't copy content)
• •Focus on DECISION + RATIONALE only

❌ BAD (Too verbose):

## [2026-02-01] Auth System
- **Goal:** User wanted login system with high security and good performance...
- **Decision:** Implemented JWT with HttpOnly cookies following OWASP guidelines...
- **Rationale:** According to Client_Side_Security_Architecture.md section 3.2...

✅ GOOD (Token-efficient):

## Auth: JWT + HttpOnly Cookies
Why: XSS protection (ref: Client_Side_Security)
Deps: UserDB, Redis
Risk: Rate limit strict

Ultra-Compact Template

Create architecturehistory.md in project root:

## [Feature]: [Decision]
Why: [One-line rationale] (ref: KB_File)
Deps: [Component list]
Risk: [One-line debt/risk]

What to Log (Keep it SHORT)

1. •Feature:Decision - Brief identifier
2. •Why - One-line rationale + KB reference
3. •Deps - Component names only
4. •Risk - Known limitation (optional)

Token Budget per Entry

• •Maximum 100 tokens per entry
• •Maximum 50 entries per file
• •If file grows large → Archive old entries to architecturehistory.archive.md

Why Minimal Memory Matters

• •Fast context loading - Less token burn on file read
• •Quick scanning - Find relevant decisions instantly
• •Cross-session persistence - Survives context resets
• •Human-readable - Developers can understand at a glance

📚 Knowledge Base Structure

This skill references 8 comprehensive knowledge base files located in the knowledgeAI/ directory:

1. AI Guardrail Architecture

Scope: Distributed systems, security architecture, microservices
Keywords: distributed-systems, consistency, consensus, resilience, microservices, CAP-theorem, circuit-breaker, event-driven, SRE, observability, zero-trust

Key Topics:

• •Core Distributed Systems Architecture (Consistency, Consensus, Resilience)
• •Event-Driven Architecture & Async Patterns
• •Database Engineering & Evolution
• •Site Reliability Engineering (SRE) & Observability
• •Security & Compliance Architecture
• •Infrastructure as Code (IaC) & FinOps

2. Architectural Resilience in High-Integrity Systems

Scope: Reliability patterns, observability, chaos engineering
Keywords: observability, distributed-tracing, OpenTelemetry, chaos-engineering, rate-limiting, transactional-outbox, CDC, cache-stampede

Key Topics:

• •Observability Data Plane (Tracing, Metrics Hierarchy, Structured Logging)
• •Property-Based Testing & Chaos Engineering
• •High-Scale Data Access and Caching
• •Asynchronous Architecture and Event Reliability
• •Traffic Control and Rate Limiting
• •Infrastructure as Code Security

3. Client-Side Security Architecture

Scope: Web security, XSS prevention, token storage
Keywords: XSS, CSP, DOM-based-XSS, supply-chain-risk, JWT-storage, HttpOnly-cookies, BFF-pattern, DOMPurify

Key Topics:

• •Modern Client-Side Threat Landscape (Mutation XSS, Magecart)
• •Content Security Policy (Strict CSP, Nonce-Based Architecture)
• •Secure Storage Patterns for Authentication Tokens
• •HTML Sanitization and XSS Prevention
• •Supply Chain Risk Mitigation

4. Comprehensive Architectural Framework

Scope: Database concurrency, ORM performance, security anomalies
Keywords: ACID, isolation-levels, dirty-read, phantom-read, N-plus-1, ORM, optimistic-locking, Redlock, BOLA, IDOR

Key Topics:

• •Database Concurrency and Transactional Integrity
• •The N+1 Query Problem and ORM Performance
• •Distributed Locking and Coordination (Redlock, Fencing Tokens)
• •Frontend State Management and Security
• •API Security and Access Control

5. Contextual Architecture Decision Matrix

Scope: Overengineering prevention, technology selection
Keywords: overengineering, premature-optimization, YAGNI, Cynefin, ADR, right-sizing, microservices, monolith

Key Topics:

• •The Overengineering Trap Recognition
• •Right-Sizing Principles
• •Decision Framework (Cynefin Framework, Decision Checklist)
• •Technology Selection Matrix
• •Architecture Decision Records (ADR)
• •Anti-Patterns to Avoid

6. Defensive UI Engineering

Scope: Resilient UI patterns, error handling
Keywords: error-boundaries, graceful-degradation, retry-logic, skeleton-screens, optimistic-UI, a11y, ARIA-live

Key Topics:

• •Resilient UI Architecture (Error Boundaries)
• •Error Handling Patterns (Network Errors, User-Facing Messages)
• •Loading and Skeleton States
• •Optimistic UI with Rollback
• •Accessibility in Error States

7. End-to-End Type Safety

Scope: TypeScript, schema validation, runtime validation
Keywords: type-safety, TypeScript, schema-sharing, OpenAPI, tRPC, Zod, runtime-validation, contract-testing

Key Topics:

• •Type Safety Spectrum (Compile-time to Runtime)
• •Schema Sharing Strategies (Monorepo, OpenAPI, tRPC)
• •Runtime Validation (Zod Patterns, Type Guards)
• •Contract Testing (Pact Implementation)
• •Type-Safe API Clients

8. Modern Web Performance Engineering

Scope: Core Web Vitals, bundle optimization, rendering patterns
Keywords: Core-Web-Vitals, LCP, FID, CLS, INP, bundle-analysis, tree-shaking, code-splitting, SSR, SSG, hydration

Key Topics:

• •Core Web Vitals Optimization (LCP, FID, CLS, INP)
• •Bundle Analysis and Optimization
• •Tree-Shaking Strategies
• •Loading Strategies (Code Splitting, Resource Hints)
• •Rendering Patterns (SSR, SSG, CSR, Hydration)

🔍 Quick Reference: Keyword Mapping

Use this table to quickly identify which knowledge base to reference:

🏗️ Cross-Cutting Concerns

Many topics span multiple knowledge bases. Reference them together for comprehensive coverage:

💡 How to Use This Skill

When Implementing Features:

1. •Identify keywords from the user's request or task description
2. •Match keywords to the appropriate knowledge base using the Quick Reference table
3. •Reference the relevant sections to understand best practices and anti-patterns
4. •Apply the guidelines to ensure production-grade code quality
5. •Check cross-cutting concerns for related patterns

When Reviewing Code:

1. •Scan code for potential issues matching knowledge base topics
2. •Reference relevant sections for validation
3. •Identify anti-patterns and suggest improvements
4. •Ensure security, performance, and reliability standards

🎯 Core Principles

The Silent Failure Problem

AI-generated code often contains issues that pass tests but fail in production:

• •✅ What Testing Catches: Syntax errors, runtime exceptions, failing assertions
• •❌ What Testing Misses:
  • •Race conditions and concurrency anomalies
  • •Subtle security vulnerabilities (XSS variants, BOLA attacks)
  • •Performance degradation (N+1 queries, thundering herd)
  • •Architectural anti-patterns
  • •Resilience failure modes

Guardrail Philosophy

This knowledge base acts as guardrails, not rigid rules:

• •Context-aware: Right-sizing over overengineering
• •Evidence-based: Grounded in real production failures
• •AI-optimized: Structured for efficient LLM context usage
• •Pragmatic: Provides patterns, not mandates

📁 File Naming Convention

All knowledge base files follow the pattern:

{Descriptive_Name}_Knowledge_Base.md

This enables:

• •Easy identification of topics
• •Programmatic access
• •Consistent organization

🔗 Integration with Opencode

This skill automatically activates when the user mentions relevant keywords in their requests. The knowledge base provides:

• •Proactive guidance during code generation
• •Quality validation during code review
• •Best practice enforcement for production readiness
• •Cross-reference suggestions for comprehensive solutions

📖 Example Usage

Example 1: Database Concurrency

User: "Add a transaction to transfer funds between accounts"

Skill Action:
1. Detect keywords: "transaction", "database"
2. Reference: Comprehensive Architectural Framework
3. Apply: Isolation levels, deadlock prevention, ACID compliance
4. Cross-check: AI Guardrail Architecture for distributed patterns

Example 2: API Security

User: "Store JWT tokens in the frontend"

Skill Action:
1. Detect keywords: "JWT", "token", "frontend"
2. Reference: Client-Side Security Architecture
3. Apply: Secure storage patterns, HttpOnly cookies, BFF pattern
4. Warn: localStorage risks, in-memory storage alternatives

Example 3: Type Safety

User: "Validate API response data"

Skill Action:
1. Detect keywords: "validate", "API", "TypeScript"
2. Reference: End-to-End Type Safety
3. Apply: Zod runtime validation, type guards
4. Suggest: Schema sharing between frontend and backend

🎓 Target Audience

• •AI Coding Assistants: Claude, GPT-4, Copilot, etc.
• •Senior Engineers: Architecture review reference
• •Mid-level Developers: Learning advanced patterns
• •Tech Leads: Team standards and guidelines

🙏 Sources

This knowledge base synthesizes:

• •Production incident post-mortems
• •Industry best practices (Google SRE, OWASP, W3C)
• •Academic distributed systems research
• •Real-world AI coding assistant failure modes

Remember: The goal is not to prevent all failures, but to prevent the failures that matter in production.