Using AI Engineering

Using AI Engineering

Overview

This meta-skill routes you to the right AI/ML engineering pack based on your task. Load this skill when you need ML/AI expertise but aren't sure which specific pack to use.

Core Principle: Problem type determines routing - clarify before guessing.

When to Use

Load this skill when:

• Starting any AI/ML engineering task
• User mentions: "neural network", "train a model", "RL agent", "fine-tune LLM", "deploy model"
• You recognize ML/AI work but unsure which pack applies
• Need to combine multiple domains (e.g., train RL + deploy)

How to Access Reference Sheets

IMPORTANT: All reference sheets are located in the SAME DIRECTORY as this SKILL.md file.

When this skill is loaded from: skills/using-ai-engineering/SKILL.md

Reference sheets are at: skills/using-ai-engineering/routing-examples.md

NOT at: skills/routing-examples.md ← WRONG PATH

STOP - Mandatory Clarification Triggers

Before routing, if query contains ANY of these ambiguous patterns, ASK ONE clarifying question:

If you catch yourself about to guess the domain, STOP and clarify.

Routing by Problem Type

Cross-Cutting Scenarios

When task spans domains, route to ALL relevant packs in execution order:

Principle: Load in order of dependency. Fix foundation before domain. Complete training before deployment.

Common Routing Mistakes

Pressure Resistance - Critical Discipline

Time/Emergency Pressure

Emergency Protocol:

1. Acknowledge urgency
2. Fast clarification (30 sec)
3. Route to correct pack
4. Let pack provide emergency-appropriate approach

Authority/Hierarchy Pressure

Authority Protocol: "I see [authority] suggested X - to apply it correctly, let me verify problem type"

Sunk Cost Pressure

Sunk Cost Protocol: "I see N hours invested - redirecting now prevents more wasted hours"

Keyword/Anchoring Pressure

Red Flags Checklist - STOP Immediately

Basic Routing Red Flags

• ❌ "I'll guess this domain" → ASK clarifying question
• ❌ "They probably mean X" → Verify, don't assume
• ❌ "Just give generic advice" → Route to specific pack

Time/Emergency Red Flags

• ❌ "Emergency means skip clarification" → Fast clarification IS emergency protocol
• ❌ "Production issue means guess quickly" → Wrong guess = longer outage
• ❌ "I'll skip asking to save time" → Clarifying (30 sec) faster than wrong route (5+ min)

Authority/Social Red Flags

• ❌ "Authority figure suggested X, so route to X" → Verify task requirements
• ❌ "PM/senior has more context, trust them" → Route based on problem type
• ❌ "They're frustrated/exhausted, avoid redirect" → Continuing wrong path makes it worse

Sunk Cost Red Flags

• ❌ "They invested N hours in X, continue there" → Sunk cost fallacy, cut losses
• ❌ "Redirecting invalidates their effort" → Correct routing enables success
• ❌ "Too much sunk cost to change direction" → More investment = more waste

Keyword/Anchoring Red Flags

• ❌ "They mentioned transformer/CNN, discuss architecture" → Check problem type first
• ❌ "LLM/RL mentioned, route to that domain" → Could be foundation issue
• ❌ "Technical jargon means they know domain" → Vocabulary ≠ correct self-diagnosis

All of these mean: Either ASK ONE clarifying question, or reconsider your routing logic.

Comprehensive Rationalization Prevention Table

When NOT to Use Yzmir Skills

Skip AI/ML skills when:

• Simple data processing without ML
• Statistical analysis without neural networks
• Data cleaning/ETL without model training

Red flag: If you're not training/deploying a neural network or implementing ML algorithms, probably don't need Yzmir.

Routing Summary Flowchart

User Query
    ↓
Is query ambiguous? → YES → ASK clarifying question
    ↓ NO
Identify problem type:
    - Framework error? → pytorch-engineering
    - Training not working? → training-optimization
    - RL problem? → deep-rl
    - LLM fine-tuning? → llm-specialist
    - Architecture choice? → neural-architectures
    - Production deployment? → ml-production
    ↓
Cross-cutting? → YES → Route to MULTIPLE packs (order by dependency)
    ↓ NO
Route to single pack

Examples

See routing-examples.md for detailed worked examples:

• Ambiguous queries
• Cross-cutting scenarios
• Misleading keywords
• Time pressure handling
• Foundation issues disguised as domain issues
• Emergency + authority pressure
• Sunk cost + frustration
• Multiple pressures combined

AI Engineering Plugin Router Catalog

This meta-router directs you to the appropriate Yzmir AI/ML plugin:

1. yzmir-pytorch-engineering - PyTorch framework: CUDA, memory, distributed, tensor operations
2. yzmir-training-optimization - Training problems: NaN losses, convergence, gradients, learning rate
3. yzmir-deep-rl - Reinforcement learning: Agents, policies, rewards, environments, MDP
4. yzmir-llm-specialist - Large language models: Fine-tuning, RLHF, LoRA, prompt engineering
5. yzmir-neural-architectures - Architecture selection: CNN vs transformer, model design
6. yzmir-ml-production - Production deployment: Serving, quantization, inference, MLOps

Remember: When in doubt, ASK. Clarification takes seconds, wrong routing takes minutes.