Menú
Skills en este repositorio
SkillsMP ha recopilado 606 skills de merceralex397-collab/meta-skill-engineering. Abre una skill para revisar su origen y sus detalles.
merceralex397-collab/meta-skill-engineeringMostrando 40 de 606 skills recopiladas.
Map a repository's directory roles so the agent can distinguish canonical source, generated output, operational scripts, tests, archives, and user-owned areas before making changes. Use when the repo is large, layered, scaffolded, or has multiple output…
Idioma del texto original: inglés
Delegate read-only evidence gathering to scoped sub-agents with explicit output contracts, then synthesize findings into persistent artifacts. Use when the user says "research task", "investigate codebase", "gather evidence", "read-only exploration",…
Idioma del texto original: inglés
Synthesize research from multiple sources into structured summaries — extract key findings, compare methodologies, identify gaps, organize by theme, and produce literature reviews with proper citations. Use when reviewing academic papers, comparing technical…
Idioma del texto original: inglés
Orchestrates a team of specialist agents to deliver a milestone through work breakdown, assignment, and progress tracking. Trigger — "coordinate the agents", "break down this milestone", "assign work to specialists", "lead the team on this feature",…
Idioma del texto original: inglés
Execute a test scenario against a team, observing coordination pattern behaviors, evaluating acceptance criteria, and generating a structured RESULT.md. Use when validating that a team's coordination pattern produces the expected behaviors during a realistic…
Idioma del texto original: inglés
Design and implement LLM tool-use and agent workflows — register tool schemas, build agent control loops, parse observations, handle multi-step reasoning chains, and manage tool errors and retries. Use when building agents that call external tools/APIs,…
Idioma del texto original: inglés
Trigger — "verify before moving on", "add a gate check", "require proof before next step", "evidence checklist", "quality gate", "don't proceed until tests pass". Skip — task is exploratory with no defined stages; user explicitly says "skip checks"; the…
Idioma del texto original: inglés
Validate deliverables and build evidence trails when work passes between agents. Covers expected outcome specification before execution, structured evidence generation during execution, deliverable validation against external anchors after execution, fidelity…
Idioma del texto original: inglés
Trigger — "save workflow state", "persist progress between sessions", "checkpoint the current run", "resume where we left off", "state file", "serialize workflow progress". Skip — task completes in a single session with no resumption needed; state is already…
Idioma del texto original: inglés
Guides full and parameter-efficient fine-tuning (LoRA, QLoRA) of LLMs using PEFT, TRL SFTTrainer, and BitsAndBytes. Covers adapter config, hyperparameter selection, data formatting, and evaluation. Use when adapting a pretrained model to a specific task or…
Idioma del texto original: inglés
Compile and run llama.cpp for local inference — GGUF quantization, context sizing, and GPU offloading. Use when building llama.cpp from source, converting models to GGUF, configuring n_gpu_layers for partial offload, or tuning context size and batch…
Idioma del texto original: inglés
Plans and executes end-to-end LLM creation from architecture design through pretraining, instruction tuning, and alignment. Covers scaling laws, compute budgets, tokenizer training, distributed training infrastructure, and evaluation checkpoints. Use when…
Idioma del texto original: inglés
Creates task-specific evaluations for model quality, latency, reliability, and regressions. Use this when the work involves models, inference, training, evaluation, or LLM system design or a task in the "AI / LLM Runtime and Integration Skills" family needs…
Idioma del texto original: inglés
Covers local model runtime selection, deployment, constraints, and practical tradeoffs. Use this when the work involves models, inference, training, evaluation, or LLM system design or a task in the "AI / LLM Runtime and Integration Skills" family needs…
Idioma del texto original: inglés
Design and implement transformer model architectures including attention variants (MHA, GQA, MQA), positional encodings (RoPE, ALiBi), normalization strategies (Pre-LN, RMSNorm), and FFN activations (SwiGLU, GELU). Use when defining or modifying LlamaConfig,…
Idioma del texto original: inglés
Merge multiple fine-tuned LLM checkpoints using mergekit with methods like linear interpolation, SLERP, TIES, DARE, task arithmetic, and frankenmerging. Use when combining specialized model capabilities without retraining — e.g., merging a code model with a…
Idioma del texto original: inglés
Route LLM requests to different models based on complexity, cost, and latency requirements. Use when implementing a model router that dispatches to small/large models, designing fallback chains, or optimizing cost by routing simple queries to cheaper models.…
Idioma del texto original: inglés
Choose the right LLM for a task based on capability, cost, latency, and context window. Use when evaluating models for a new project, comparing model benchmarks, deciding between API and self-hosted, or sizing context window requirements. Do not use for…
Idioma del texto original: inglés
Design Mixture-of-Experts transformer architectures including router/gating design (top-k softmax), expert FFN configuration, load balancing loss, capacity factors, and expert parallelism. Use when implementing MixtralConfig, Switch Transformer, or custom MoE…
Idioma del texto original: inglés
Integrate vision, audio, and multimodal LLM capabilities into applications. Use when sending images to vision models, processing audio with Whisper, combining text+image inputs, or building multimodal pipelines. Do not use for text-only LLM tasks or…
Idioma del texto original: inglés
Optimize LLM inference for CPU-only environments — quantization, threading, and memory mapping. Use when running models without GPU, optimizing llama.cpp for CPU, choosing quantization for RAM-constrained systems, or deploying inference on commodity hardware.…
Idioma del texto original: inglés
Configure, serve, and manage local LLMs with Ollama — write Modelfiles, pull/push models, set GPU layers and context windows, call chat/generate/embeddings API endpoints, and troubleshoot serving issues. Use when a task involves ollama serve, ollama run,…
Idioma del texto original: inglés
Applies post-training quantization (GPTQ, AWQ, GGUF) and quantization-aware training to reduce LLM memory footprint and inference cost. Use when the task involves bit-width selection, calibration, weight quantization, or evaluating perplexity degradation from…
Idioma del texto original: inglés
Select and apply model quantization formats (GGUF, GPTQ, AWQ, bitsandbytes) with appropriate bit widths, calibration data, and quality-latency tradeoffs. Use when choosing quantization format for deployment, converting models between formats, tuning…
Idioma del texto original: inglés
Evaluate and improve retrieval system quality using NDCG, MRR, recall@k, precision@k, and hit-rate metrics with relevance-annotated query sets. Use when measuring retrieval pipeline accuracy, comparing retrieval configurations, building evaluation datasets,…
Idioma del texto original: inglés
Trains and evaluates reward models that score LLM outputs for RLHF pipelines using Bradley-Terry preference modeling and TRL's RewardTrainer. Use when building a scalar reward signal from pairwise human preferences, diagnosing reward hacking, or choosing…
Idioma del texto original: inglés
Implements red teaming, refusal training, Constitutional AI, and safety RLHF to align LLMs against harmful outputs while preserving helpfulness. Use when designing safety data pipelines, evaluating jailbreak robustness with benchmarks like…
Idioma del texto original: inglés
Implement input and output guardrails for LLM applications — add prompt injection detection, PII scrubbing, toxicity filtering, content classification, output validation, and refusal handling. Use when hardening an LLM-powered feature against adversarial…
Idioma del texto original: inglés
Designs and deploys LLM inference infrastructure using vLLM, TGI, or TensorRT-LLM with continuous batching, PagedAttention KV cache management, and speculative decoding. Use when configuring serving frameworks, optimizing throughput/latency, setting up…
Idioma del texto original: inglés
Design, train, and evaluate tokenizers (BPE, SentencePiece unigram) for LLMs. Use when selecting vocabulary size, defining special tokens, training a tokenizer on a corpus, analyzing fertility/compression, or handling multilingual coverage. Covers the…
Idioma del texto original: inglés
Transcribe audio files to text with optional diarization and known-speaker hints. Use when a user asks to transcribe speech from audio/video, extract text from recordings, or label speakers in interviews or meetings.
Idioma del texto original: inglés
Handles vLLM-based serving, batching, throughput, and API integration where GPU or server setups justify it. Use this when the work involves models, inference, training, evaluation, or LLM system design or a task in the "AI / LLM Runtime and Integration…
Idioma del texto original: inglés
Orchestrate end-to-end machine learning pipelines using Prefect or Airflow with DAG construction, task dependencies, retry logic, scheduling, monitoring, and integration with MLflow, DVC, and feature stores for production ML workflows. Use when automating…
Idioma del texto original: inglés
A broader external seed for Python-centric LLM and ML repository work. Use this when the work involves models, inference, training, evaluation, or LLM system design or a task in the "AI / LLM Runtime and Integration Skills" family needs repeatable procedure…
Idioma del texto original: inglés
Generates synthetic instruction-response training data using Self-Instruct, Evol-Instruct, and distillation from strong models via OpenAI/Anthropic APIs or vLLM batch inference. Use when creating instruction-tuning datasets, evolving seed tasks for…
Idioma del texto original: inglés
Configure distributed LLM training infrastructure—DDP, FSDP, DeepSpeed ZeRO, multi-node orchestration, checkpointing, fault tolerance, and mixed precision. Use when setting up torchrun/accelerate/deepspeed jobs, writing SLURM scripts, tuning NCCL, or…
Idioma del texto original: inglés
Review a data analysis for quality, correctness, and reproducibility. Covers data quality assessment, assumption checking, model validation, data leakage detection, and reproducibility verification. Use when reviewing a colleague's analysis before…
Idioma del texto original: inglés
Conduct a peer review of research methodology, experimental design, and manuscript quality. Covers methodology evaluation, statistical appropriateness, reproducibility assessment, bias identification, and constructive feedback. Use when reviewing a…
Idioma del texto original: inglés
Build a custom MCP (Model Context Protocol) server that exposes domain-specific tools to AI assistants. Covers server implementation in Node.js or R, tool definitions, transport configuration, and testing with Claude Code. Use when you need to expose custom…
Idioma del texto original: inglés
Configure the putior MCP server to expose 16 workflow visualization tools to AI assistants. Covers Claude Code and Claude Desktop setup, dependency installation (mcptools, ellmer), tool verification, and optional ACP server configuration for agent-to-agent…
Idioma del texto original: inglés