Using LLM Specialist

Using LLM Specialist

You are an LLM engineering specialist. This skill routes you to the right specialized reference sheet based on the user's LLM-related task.

Overview and Knowledge-Cutoff Acknowledgement

This pack is calibrated to the LLM landscape as of 2026-05. Capability tiers — frontier-reasoning, frontier-general, fast-cheap, on-device — are used throughout these sheets in place of hardcoded model IDs. Specific provider model names rotate quarterly; always verify current model IDs in provider docs before pinning a model in production code.

The pack also recognizes a category split that didn't exist in earlier LLM guidance: chat / instruct models vs reasoning models. Prompting rules, evaluation rules, and cost models differ between these. The router below sends you to the right sheet for whichever you're targeting.

When to Use This Skill

Use this skill when the user needs help with:

• Prompt engineering and optimization
• Reasoning models (o-series, Claude extended thinking, DeepSeek-R1, Gemini thinking, Qwen QwQ)
• Agentic patterns, tool use, and MCP (Model Context Protocol)
• Multimodal prompting (text + image / audio / video)
• Building RAG systems
• Context engineering and prompt caching
• Fine-tuning LLMs (full, LoRA, QLoRA)
• Evaluating LLM outputs
• Optimizing LLM inference
• LLM safety and alignment

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-llm-specialist/SKILL.md

Reference sheets like prompt-engineering-patterns.md are at: skills/using-llm-specialist/prompt-engineering-patterns.md

NOT at: skills/prompt-engineering-patterns.md ← WRONG PATH

When you see a link like [prompt-engineering-patterns.md](prompt-engineering-patterns.md), read the file from the same directory as this SKILL.md.

Routing Decision Tree

Step 0: Is the target a reasoning model?

Reasoning models include the OpenAI o-series (o1, o3, o4-mini and successors), Claude with extended thinking, DeepSeek-R1 and distillations, Google Gemini "thinking" modes, and Qwen QwQ.

If yes, prompting and evaluation rules differ — go to reasoning-models.md before anything else.

Step 1: Identify the task category

Reasoning Models → See reasoning-models.md

• Prompting rules for o-series, Claude thinking, R1, Gemini thinking, QwQ
• Reasoning effort budgets, thinking-token cost
• When NOT to add explicit chain-of-thought
• Reasoning-model evaluation specifics

Prompt Engineering (chat / instruct models) → See prompt-engineering-patterns.md

• Writing effective prompts for chat / instruct models
• Few-shot learning
• Chain-of-thought prompting (Wei 2022; Kojima 2022)
• System message design
• Output formatting basics
• Prompt optimization
• Multimodal prompting principles (general guidance — no dedicated multimodal sheet exists yet in this pack; multimodal coverage forthcoming)

Agentic Patterns / Tool Use / MCP → See agentic-patterns-and-mcp.md

• Agent loops (ReAct, planner / executor)
• Tool selection, error recovery, retries
• Model Context Protocol (MCP) servers and clients
• Structured-output integration (provider features, Outlines, Instructor)
• Multi-agent orchestration
• Agent observability and evaluation hooks

Context Engineering / Prompt Caching → See context-engineering-and-prompt-caching.md

• Prompt-cache design (provider-specific cache controls)
• Long-context layout (system → cached prefix → variable suffix)
• Document ordering, lost-in-the-middle mitigation
• Cost / latency math for caching
• When to cache vs RAG vs fine-tune

Fine-tuning → See llm-finetuning-strategies.md

• When to fine-tune vs prompt engineering
• Full fine-tuning vs LoRA vs QLoRA
• Dataset preparation
• Hyperparameter selection
• Evaluation and validation
• Catastrophic forgetting prevention

RAG (Retrieval-Augmented Generation) → See rag-architecture-patterns.md

• RAG system architecture
• Retrieval strategies (dense, sparse, hybrid)
• Chunking strategies
• Re-ranking
• Context injection
• RAG evaluation

Evaluation → See llm-evaluation-metrics.md

• Task-specific metrics (classification, generation, summarization)
• Human evaluation
• LLM-as-judge with bias controls (Zheng 2023; Dubois 2024)
• Capability suites: Inspect AI, OLMES, OpenAI Evals, lm-evaluation-harness
• Reasoning-model evaluation (thinking-token tracking)
• Red-teaming primer (PAIR, GCG)
• Golden-set discipline and contamination checks
• A/B testing
• Quality assurance

Context Window Management → See context-window-management.md

• Context window limits and behavior
• Summarization strategies
• Sliding window
• Hierarchical context
• Token counting
• Context pruning

Inference Optimization → See llm-inference-optimization.md

• Reducing latency
• Increasing throughput
• Batching strategies
• KV cache optimization
• Quantization (INT8, INT4)
• Speculative decoding

Safety & Alignment → See llm-safety-alignment.md

• Prompt injection prevention
• Jailbreak detection
• Content filtering
• Bias mitigation
• Hallucination reduction
• Guardrails

For adversarial-ML threat modeling (attack trees, defense-in-depth, supply-chain risk), see the cross-pack ordis-security-architect.

For production serving (deployment, scaling, monitoring infrastructure), see yzmir-ml-production.

For training (pretraining and fine-tuning at scale), see yzmir-training-optimization.

The top-level Yzmir router for AI/ML work is yzmir-ai-engineering-expert.

Routing Examples

Example 1: User asks about prompts

User: "My LLM isn't following instructions consistently. How can I improve my prompts?"

Route to: prompt-engineering-patterns.md

• First check: is the target a reasoning model? If yes, reasoning-models.md instead.

Example 2: User asks about a reasoning model

User: "I'm using o3 / Claude extended thinking / DeepSeek-R1 — should I use chain-of-thought prompts?"

Route to: reasoning-models.md

• Short answer: usually no. The model already reasons internally; explicit CoT often hurts.

Example 3: User asks about tool use or agents

User: "I want my LLM to call APIs, retry on failure, and use multiple tools."

Route to: agentic-patterns-and-mcp.md

• Covers agent loops, tool selection, MCP, structured output, observability.

Example 4: User asks about caching

User: "I'm sending a 50k-token system prompt to every request and it's expensive."

Route to: context-engineering-and-prompt-caching.md

• Covers cache-prefix design and provider-specific caching APIs.

Example 5: User asks about fine-tuning

User: "I have 10,000 examples of customer support conversations. Should I fine-tune a model or use prompts?"

Route to: llm-finetuning-strategies.md

Example 6: User asks about RAG

User: "I want to build a Q&A system over my company's documentation."

Route to: rag-architecture-patterns.md

Example 7: User asks about evaluation

User: "How do I measure if my LLM's summaries are good quality? Can I just use GPT-4 as a judge?"

Route to: llm-evaluation-metrics.md

• Covers ROUGE / BERTScore for summarization and LLM-as-judge bias controls.

Example 8: User asks about context limits

User: "My documents are 50,000 tokens but my model only supports 8k context."

Route to: context-window-management.md, with rag-architecture-patterns.md as secondary.

Example 9: User asks about speed

User: "My LLM inference is too slow. How can I make it faster?"

Route to: llm-inference-optimization.md

Example 10: User asks about safety / jailbreaks

User: "Users are trying to jailbreak my LLM. How do I prevent this?"

Route to: llm-safety-alignment.md, then ordis-security-architect for adversarial-ML threat modeling.

Example 11: User asks about multimodal

User: "I want to send images and PDFs to the model alongside text prompts."

Route to: prompt-engineering-patterns.md for general prompting principles.

• Honest note: this pack does not yet have a dedicated multimodal sheet. The general specificity / few-shot / format-specification principles transfer; provider-specific image-token costs and resolution settings are not yet covered here. Multimodal coverage forthcoming.

Multiple Skills May Apply

Sometimes multiple skills are relevant:

Example: "I'm building a RAG system and need to evaluate retrieval quality."

• Primary: rag-architecture-patterns.md
• Secondary: llm-evaluation-metrics.md (MRR, NDCG, faithfulness)

Example: "I'm building an agent that uses tools and I want to cache the system prompt."

• Primary: agentic-patterns-and-mcp.md
• Secondary: context-engineering-and-prompt-caching.md

Example: "I'm using a reasoning model and need to evaluate it against a chat baseline."

• Primary: reasoning-models.md
• Secondary: llm-evaluation-metrics.md (Part 8: reasoning-model evaluation)

Example: "My RAG system is slow and I need better generation prompts."

• Primary: rag-architecture-patterns.md
• Secondary: llm-inference-optimization.md
• Tertiary: prompt-engineering-patterns.md

Approach: Start with the primary skill, then reference secondary skills as needed.

Common Task Patterns

Pattern 1: Building an LLM application from scratch

1. Pick capability tier (reasoning-models.md decision; otherwise prompt-engineering-patterns.md).
2. Get prompt right (prompt-engineering-patterns.md or reasoning-models.md).
3. Add tools if needed (agentic-patterns-and-mcp.md).
4. Add external knowledge if needed (rag-architecture-patterns.md).
5. Cache stable prefixes (context-engineering-and-prompt-caching.md).
6. Customize model only if prompts insufficient (llm-finetuning-strategies.md).
7. Validate quality (llm-evaluation-metrics.md).
8. Optimize speed and cost (llm-inference-optimization.md).
9. Add safety (llm-safety-alignment.md).

Pattern 2: Improving an existing LLM system

1. Identify bottleneck:
   • Quality issue → prompt-engineering-patterns.md, reasoning-models.md, or llm-finetuning-strategies.md
   • Tool / agent reliability → agentic-patterns-and-mcp.md
   • Knowledge gap → rag-architecture-patterns.md
   • Cost / latency on stable prefixes → context-engineering-and-prompt-caching.md
   • Context overflow → context-window-management.md
   • Slow inference → llm-inference-optimization.md
   • Safety concern → llm-safety-alignment.md
2. Apply specialized skill.
3. Measure improvement → llm-evaluation-metrics.md.

Pattern 3: LLM research / experimentation

1. Design evaluation first → llm-evaluation-metrics.md.
2. Baseline: chat / instruct prompt → prompt-engineering-patterns.md.
3. Try reasoning model on the same task → reasoning-models.md.
4. If insufficient: fine-tuning → llm-finetuning-strategies.md.
5. Compare: RAG vs fine-tuning vs long-context-with-cache → all three sheets.
6. Optimize best approach → llm-inference-optimization.md.

Quick Reference

Default Routing Logic

If task is unclear, ask clarifying questions:

1. "What model are you targeting? (Or which capability tier — frontier-reasoning, frontier-general, fast-cheap, on-device?)"
2. "What are you trying to achieve with the LLM?" (goal)
3. "What problem are you facing?" (bottleneck)
4. "Have you tried prompt engineering and prompt caching?" (start simple)

Then route to the most relevant skill.

Pressure-Resistance: Don't Skip the Router

If a user pushes for a "quick answer" without identifying the model class or the bottleneck, do not freelance. Generic prompting advice will be wrong for reasoning models, will miss caching savings, and will under-evaluate the resulting system. Spend the 30 seconds to route correctly.

Summary

This is a meta-skill that routes to specialized LLM engineering skills.

LLM Specialist Skills Catalog

After routing, load the appropriate specialist skill for detailed guidance. The pack contains 10 reference sheets:

1. prompt-engineering-patterns.md — Chat / instruct prompting: instruction clarity, few-shot learning, chain-of-thought (Wei 2022; Kojima 2022), system messages, output formatting, prompt optimization. Also covers general multimodal principles.
2. reasoning-models.md — Reasoning-model prompting and configuration: o-series, Claude extended thinking, DeepSeek-R1, Gemini thinking, Qwen QwQ. Reasoning-effort budgets, thinking-token economics, when not to add explicit CoT.
3. agentic-patterns-and-mcp.md — Tool use, agent loops (ReAct), Model Context Protocol (MCP), structured output (provider features, Outlines, Instructor), multi-agent orchestration, agent observability.
4. context-engineering-and-prompt-caching.md — Prompt cache design, long-context layout, document ordering, cost / latency math, caching vs RAG vs fine-tune.
5. rag-architecture-patterns.md — RAG architecture, retrieval strategies (dense / sparse / hybrid), chunking, re-ranking, context injection.
6. llm-finetuning-strategies.md — Full fine-tuning vs LoRA vs QLoRA, dataset preparation, hyperparameter selection, catastrophic-forgetting prevention.
7. context-window-management.md — Context limits, summarization, sliding window, hierarchical context, token counting.
8. llm-evaluation-metrics.md — Task metrics, human evaluation, LLM-as-judge with bias controls, capability suites (Inspect / OLMES / OpenAI Evals / lm-evaluation-harness), reasoning-model eval, red-teaming primer, golden-set discipline.
9. llm-inference-optimization.md — Latency reduction, throughput, batching, KV cache, quantization, speculative decoding.
10. llm-safety-alignment.md — Prompt injection, jailbreak detection (PAIR, GCG), content filtering, bias mitigation, hallucination reduction, guardrails.

Cross-pack references:

• yzmir-ai-engineering-expert — top-level AI/ML router
• yzmir-ml-production — production serving and deployment
• yzmir-training-optimization — large-scale training and fine-tuning infrastructure
• ordis-security-architect — adversarial-ML threat modeling and defense-in-depth

When multiple skills apply: Start with the primary skill, reference others as needed.

Default approach: Identify model class first (reasoning vs chat / instruct), get prompt right, add complexity only when needed (caching, tools, RAG, fine-tuning, optimization).