AI & LLM Engineering skills for Codex
Skills for building with models rather than just calling them: the OpenAI Responses API, Agents SDK, Realtime voice and Structured Outputs, RAG pipelines with chunking, hybrid search and evaluation, agent tool design and memory, prompt-injection defenses, evals, fine-tuning, and running local models with Ollama, vLLM and llama.cpp.
curl -fsSL codexskills.dev/install.sh | sh -s -- --category ai Installs all 53 into ~/.agents/skills/. Or download the zip.
All 53 AI & LLM Engineering skills
Each opens the full SKILL.md, example prompts and every way to install it.
-
Agent Memory
Remember what matters, forget on request -
Agent Tool Design
Tools a model picks right and calls right first time -
AI Agent Evaluation
Judge agents by end state, reliability and cost -
AI Chat Interface
Chat UIs that stream, stop and recover cleanly -
AI Content Moderation
Policy first, thresholds second, humans in the loop -
Assistants to Responses
Get off the shut-down Assistants API safely -
YOLO Object Detection
Custom detectors that survive the edge device -
Document AI and OCR
Route each page, keep confidences, review the doubtful -
Feature Engineering
Features that only know the past -
Gemini API
Gemini integrations on the current SDK and API -
Hugging Face Transformers
Load, prompt and batch Hub models without surprises -
Hybrid Search and Reranking
BM25 plus vectors, fused by rank and proven on qrels -
LangChain Apps
LangChain 1.x chains and agents that stay testable -
LangGraph Agents
State graphs that pause, resume and survive crashes -
llama.cpp and GGUF
Pick the quant by perplexity, not folklore -
LlamaIndex RAG
Incremental ingestion and answers that show sources -
LLM Classification
Labels you can trust, with confidence that means something -
LLM Evals
Ship prompt changes on numbers, not vibes -
LLM Observability
Trace any answer back to prompt, context and cost -
LLM Streaming
Tokens that arrive live and stop when asked -
LoRA Fine-Tuning
QLoRA a 7B model on one GPU and prove it helped -
ML Experiment Tracking
Every number traceable to one reproducible run -
ML Model Deployment
From notebook pickle to a model API you can roll back -
Model Quantization
Smaller, faster models with the quality loss measured -
Multi-Agent Orchestration
More agents only when the eval says it pays -
Ollama Local LLMs
Local models with the context you think you have -
OpenAI Agents SDK
Few agents, bounded runs, guardrails that block -
OpenAI Batch API
Half-price bulk inference that rejoins cleanly -
OpenAI Embeddings
One vector space, measured, never mixed -
OpenAI Fine-Tuning
Fine-tune GPT-4.1 only when evals prove it pays -
OpenAI Function Calling
Tool loops that finish and fail gracefully -
OpenAI Image Generation
Generate, mask, iterate and store images -
OpenAI Realtime Voice
Voice agents that know when to stop talking -
OpenAI Responses API
Typed output items, deliberate state, clean streams -
OpenAI Structured Outputs
Typed JSON from the model, every time -
pgvector Search
HNSW that uses the index and returns a full page -
Prompt Caching and Cost
Lower the LLM bill without lowering quality -
Prompt Engineering
Versioned prompts, changed one eval at a time -
Prompt Injection Defense
Assume the text is hostile; make it not matter -
PyTorch Training
Overfit one batch, then scale to eight GPUs -
RAG Chunking
Chunks that follow structure and stand alone -
RAG Evaluation
Know whether retrieval or generation failed -
RAG Pipelines
Grounded answers with citations that resolve -
scikit-learn Pipelines
Cross-validation scores you can believe -
Semantic Caching
Cut LLM spend without serving wrong answers -
Speech to Text
Transcripts with speakers, timestamps and a WER -
Structured Extraction
JSON from messy text, with evidence and per-field scores -
Synthetic Data Generation
Test and eval data with coverage you chose -
Text to Speech
Natural narration with names pronounced right -
Time-Series Forecasting
Beat seasonal naive in a backtest, or do not ship -
Vector Databases
Pick the store from numbers, then plan the re-embed -
Vercel AI SDK
Streaming chat and tools that match your ai version -
vLLM Serving
Size KV cache from traffic, then benchmark it
No AI & LLM Engineering skill matches that.
Where to start
Install the ones that match requests you make every week. Name one with $assistants-to-responses-migration when you want it for sure; otherwise Codex picks by description.
Start with
- Assistants to ResponsesGet off the shut-down Assistants API safely
- LoRA Fine-TuningQLoRA a 7B model on one GPU and prove it helped
- OpenAI Responses APITyped output items, deliberate state, clean streams
Install a few by name
curl -fsSL codexskills.dev/install.sh | sh -s -- agent-memory agent-tool-design 104-line POSIX script. No sudo, no telemetry. Read install.sh
AI & LLM Engineering FAQ
How do I install every AI & LLM Engineering skill at once?
Run curl -fsSL https://codexskills.dev/install.sh | sh -s -- --category ai to put all 53 in ~/.agents/skills/, or add --project to install them into the current repository's .agents/skills/. You can also download the category zip and unpack it yourself.
Does installing 53 skills slow Codex down?
Hardly. At startup Codex sees only each skill's name and description, a few dozen words apiece, and reads a full SKILL.md only when a task needs it. Install the whole category if the work fits; unused skills cost almost nothing.
Which AI & LLM Engineering skill should I start with?
Start with Assistants to Responses, LoRA Fine-Tuning, OpenAI Responses API. They cover the requests that come up most in this category. Mention one with $assistants-to-responses-migration to try it.