AI & ML

serialexperimentslainnnn Skill

Model Context Protocol server and client engineering. Use when building or auditing an MCP server or client with the official SDKs (modelcontextprotocol typescript-sdk, python-sdk, go-sdk, csharp-sdk, FastMCP), declaring tools/resources/prompts primitives, choosing stdio versus Streamable HTTP transport, protocol-revision fields (server/discover, _meta, Mcp-Method/Mcp-Name headers, resultType, MRTR inputRequests, ttlMs/cacheScope), mcp.json / .mcp.json / claude_desktop_config.json server entries, server.json and the MCP registry, MCP authorization (RFC 9728 protected resource metadata, RFC 8707 resource indicators, Client ID Metadata Documents, PKCE S256), or triaging tool poisoning, rug pull, server shadowing, token passthrough and confused-deputy exposure in installed servers.

serialexperimentslainnnn Skill

Natural language processing as a discipline, deciding between a regex, a small specialised model and an LLM. Use when building text classification, named entity recognition, sentiment analysis, summarisation, semantic similarity, clustering or translation, working with spaCy (en_core_web_sm, es_core_news_sm, nlp.pipe, Doc, Span, EntityRuler), NLTK, Stanza, Flair, gensim, scikit-learn TfidfVectorizer, setfit or a fine-tuned encoder, choosing an embedding model (sentence-transformers, BAAI/bge, intfloat/e5, nomic-embed, Alibaba gte, Qwen3-Embedding, jina-embeddings) and reading its weights licence, planning a re-index after changing embedding model, evaluating with MTEB, tokenising with tiktoken, sentencepiece or huggingface tokenizers and explaining why token count is not word count, normalising Unicode with NFC/NFKC, detecting language with fastText lid, lingua or langdetect, handling accents, casing, emoji and zero-width characters, measuring macro-F1 and per-class recall on an imbalanced label set, auditing

serialexperimentslainnnn Skill

Retrieval-augmented generation treated as a retrieval problem. Use when deciding RAG versus long context versus fine-tuning, parsing and chunking documents for indexing (PDF tables, multi-column, scans, overlap, structure-aware splits), picking an embedding model and dimensionality and paying the reindex cost of changing it, running pgvector versus Qdrant/Weaviate/Milvus/Chroma/LanceDB, tuning HNSW or IVFFlat parameters (m, ef_construction, ef_search, lists, probes), hybrid dense-plus-BM25 retrieval with RRF fusion, cross-encoder reranking, metadata filters, multi-query or HyDE expansion, grounding answers with citations and refusing when the context does not support them, measuring recall@k / MRR / nDCG separately from faithfulness, per-document access control on retrieved chunks, incremental index updates and deleting embeddings on erasure requests, or judging whether GraphRAG and agentic RAG earn their cost.

serialexperimentslainnnn Skill

LLM system evaluation standards

serialexperimentslainnnn Skill

Backup and restore mechanics — how the copy is made, where it lands and how the restore is proven. Use when choosing or operating a backup tool (restic init/backup/forget --keep-*/prune/check --read-data-subset/repository format v2, Kopia snapshot/policy set/maintenance run, BorgBackup borg create/compact/check with --append-only, borgmatic.yaml, Duplicati, rsync/rsnapshot hardlink rotation, Bacula/Bareos bconsole and Director pools, tape/LTO, Veeam in a mixed estate), designing full/incremental/differential, forever-incremental and synthetic-full chains, deduplication and compression, file-level versus block-level versus whole-image versus application-consistent snapshots, quiesce with fsfreeze and pre/post hooks around a snapshot, 3-2-1 and 3-2-1-1-0 repository topology, append-only or WORM repositories, client-side repository encryption whose passphrase never lives on the copied host, write-but-not-delete agent credentials, GFS retention with prune and garbage collection, repository check/verify, automated

serialexperimentslainnnn Skill

Applied computer vision as a data problem, not a model problem. Use when defining a vision task (classification, object detection, semantic versus instance versus panoptic segmentation, multi-object tracking, OCR, keypoint/pose estimation), building or auditing an image dataset and its label quality, inter-annotator agreement, annotating with CVAT, Label Studio, labelme, Roboflow, FiftyOne or supervision, converting between COCO JSON, YOLO .txt, Pascal VOC XML, YOLO data.yaml and instances_train.json, writing albumentations or kornia augmentation pipelines and spotting augmentation that breaks the label, choosing a detector or segmenter (Ultralytics YOLO/YOLO26, RT-DETR, RF-DETR, D-FINE, DEIM, YOLOX, Detectron2, MMDetection, SAM/SAM 2/SAM 3, Grounding DINO, DINOv2/DINOv3) and reading its weights licence before shipping, computing IoU, mAP@0.5:0.95, per-class confusion or PR curves, exporting to ONNX, TensorRT, OpenVINO, LiteRT or Core ML, matching train and serve preprocessing (resize, letterbox, BGR/RGB, nor

serialexperimentslainnnn Skill

Autonomous agent engineering, provider-agnostic. Use when choosing between an autonomous agent loop and a deterministic workflow, designing an agent's tool surface and tool descriptions, bounding the agent loop (max iterations, token and wall-clock budgets, stop criteria, unproductive-loop detection), compacting or pruning stale tool results and file-backed agent memory, orchestrating coordinator/worker subagent fan-out, building on LangGraph, CrewAI, OpenAI Agents SDK, Pydantic AI, Google ADK or Microsoft Agent Framework, human approval gates for irreversible tool calls, agent sandboxing and egress containment, end-to-end agent success rate and per-task cost tracing, or the lethal trifecta and the OWASP ASI01-ASI10 agentic risks.

serialexperimentslainnnn Bundle

Use whenever choosing an instrument inside a JetBrains IDE bound through the Claude Code Native plugin — the four MCP servers code, run, vcs and ops reached as mcp__code__run, mcp__run__run, mcp__vcs__run and mcp__ops__run through domains(), tools(domain) and run(tool, args), TOON results, the Security Guard, cards, and every tool by name (read_file, search_text, find_files, list_directory, find_symbols, definition, references, implementations, file_outline, symbol_info, hierarchy, problems, project_problems, inspect, replace_text, insert_text, create_file, write_file, undo, redo, search_replace, line_ops, rename, move_file, safe_delete, reformat, optimize_imports, open_file, active_file, editor_action, psi_tree, psi_replace, mark_add, banner_show, scratch_create, build, run_configuration, run_tests, shell, terminal_tabs, session, step, frames, values, breakpoint, git_status, git_log, git_diff, git_stage, git_commit, git_branch, git_remote, vcs_open, vcs_action, commit_action, branch_op, stash, shelve, blame,

serialexperimentslainnnn Skill

Use when serving open-weight LLMs on your own infrastructure — vllm serve, llama-server and llama.cpp, ollama serve with OLLAMA_HOST/OLLAMA_NUM_PARALLEL, SGLang, TGI or TensorRT-LLM, a self-hosted OpenAI-compatible /v1/chat/completions endpoint, .gguf and .safetensors weight files and their provenance, choosing AWQ/GPTQ/FP8/NVFP4/MXFP4/Q4_K_M quantization, sizing KV cache VRAM with --max-model-len, --gpu-memory-utilization and --tensor-parallel-size, prefill versus decode and TTFT/TPOT benchmarking under sustained load, open-weight model licences, or the break-even calculation of self-hosting versus a hosted inference API.

serialexperimentslainnnn Skill

Technical leadership as a set of decisions and artifacts, not a personality trait. Use when defining or applying an engineering career ladder and the IC-versus-manager track, assigning work by Staff+ archetype (tech lead, architect, solver, right hand), writing or reviewing an ADR and classifying a decision as a one-way or two-way door, writing a design document and running a design review with acceptance criteria, deciding who decides and at what level, budgeting technical debt as an explicit business decision, protecting a blameless postmortem under management pressure, running a 1:1 with a report-owned agenda, giving written performance feedback, being asked for individual productivity metrics such as lines of code, commits, velocity or DORA per engineer, citing SPACE or DORA in a measurement argument, sizing a team and its cognitive load in the Team Topologies sense, mapping inter-team dependencies, resolving a technical disagreement and applying disagree-and-commit, deciding whether to adopt coding agent

serialexperimentslainnnn Skill

Operating a vector search engine as a piece of infrastructure. Use when sizing RAM and disk for an ANN index, choosing binary, scalar, product or rotational/RaBitQ-style vector compression, diagnosing a selective pre-filter or post-filter that collapses recall or latency, collection snapshots and restoring an index, index build and rebuild time, cold start after restart, replication factor, shards and horizontal partitioning of collections, tombstones and graph degradation after deletes, per-collection versus per-filter tenant isolation, securing engines that ship open (Qdrant service.api_key, AUTHENTICATION_ANONYMOUS_ACCESS_ENABLED, Milvus authorizationEnabled and the default root password, ports 6333/6334, 19530, 8080, 8000), embedding inversion as a privacy risk, or non-RAG similarity workloads such as recommendation, deduplication and record linkage, image or audio search, anomaly detection and clustering.

serialexperimentslainnnn Skill

Fine-tuning an existing model's weights as the last resort, after prompting and retrieval. Use when deciding between prompt engineering, RAG and training, running full fine-tuning versus PEFT with peft LoraConfig (r, lora_alpha, target_modules="all-linear", lora_dropout), QLoRA with bitsandbytes 4-bit, or DoRA and other LoRA variants, training with trl SFTTrainer, DPOTrainer, GRPOTrainer, KTOTrainer, RLOOTrainer or RewardTrainer, Unsloth or Axolotl configs, building a chat-formatted instruction dataset with chat templates and JSONL conversations, checking eval-set contamination and synthetic-data provenance, measuring catastrophic forgetting of general capability before and after, reading the weights licence of Llama, Gemma, Qwen, Mistral or DeepSeek checkpoints and whether it is OSI-approved, merging adapters with merge_and_unload versus serving adapters at runtime, or maintaining a family of fine-tuned variants.

serialexperimentslainnnn Skill

Use when an engineering team works with coding agents day to day — deciding which task to hand to an agent and which to write by hand, AGENTS.md, CLAUDE.md, .cursorrules and .github/copilot-instructions.md repository instruction files, agent permission allowlists and settings.json approval rules, filesystem scope and which credentials the agent process can read, sandboxing a coding agent on a workstation, running an agent in CI (claude-code-action, GitHub Agentic Workflows, Copilot coding agent) and scoping its token and OIDC identity, indirect prompt injection through issues, PR comments, fetched web pages, dependency READMEs and tool output, reviewing a large plausible agent-authored diff, who is accountable for a merged agent-written change, disclosure and trailer conventions for AI-generated commits, copyright and licensing of generated code, or citing productivity evidence (METR RCT, DORA report) in a rollout decision.

serialexperimentslainnnn Skill

Defending the endpoint as a control, and measuring whether the control is actually there. Use when selecting or operating an EDR/XDR agent (Microsoft Defender for Endpoint, CrowdStrike Falcon, SentinelOne, Elastic Defend) and deciding what to demand of it in a bake-off, weighing signature antivirus against behavioural telemetry, deploying application control with App Control for Business / WDAC versus AppLocker and its MSRC servicing-criteria gap, WDAC policy in audit versus enforced mode and managed installer, Attack Surface Reduction rules, LSA protection and Credential Guard, disk encryption posture with BitLocker TPM-only versus TPM+PIN, manage-bde protectors and recovery-key escrow, UEFI Secure Boot, Measured Boot and TPM PCR attestation, endpoint patch and third-party update coverage, EDR sensors on Linux (eBPF sensor versus loadable kernel module) and on macOS (Apple Endpoint Security API ceilings), the security agent itself as attack surface and as an availability risk after the July 2024 CrowdStrike

serialexperimentslainnnn Skill

Code review as an explicit quality control, not an opinion about someone else's style. Use when reviewing or authoring a pull/merge request diff, writing a PR description or review checklist, deciding what blocks a merge versus what is a suggestion, labelling review comments (Conventional Comments, nit:, blocking:), setting review SLA and stale-PR policy, assigning reviewers or debugging a CODEOWNERS bottleneck, requiring a second or specialist reviewer for high-risk changes (data migrations, authn/authz, cryptography, concurrency, infrastructure), reviewing AI-generated or agent-authored diffs, arguing about formatting in a review, escalating a review disagreement to an ADR, choosing review metrics, or replacing async review with pair or mob programming.

serialexperimentslainnnn Skill

Bare KVM/QEMU with libvirt on standalone hosts, with no management platform above it. Use when running virsh, virt-install, virt-xml, virt-clone, virt-manager, virt-viewer, virt-sysprep, virt-customize, virt-df, guestfish or libguestfs tools, qemu-img and qemu-system-x86_64, editing domain XML under /etc/libvirt/qemu, libvirtd.conf, qemu.conf or the modular daemons (virtqemud, virtnetworkd, virtstoraged, virtnodedevd, virtproxyd), defining libvirt storage pools and volumes or virtual networks (default NAT, bridge, macvtap, open vswitch portgroup), pinning a versioned machine type (pc-q35-x.y) instead of an alias, choosing host-passthrough versus a named CPU model, OVMF/UEFI nvram and swtpm virtual TPM, hugepages, numatune and vcpupin, iothreads, qcow2 versus raw versus LVM/zvol backing with cache modes none/writeback/directsync and io=native/io_uring, discard=unmap, VFIO and IOMMU groups for PCI or GPU passthrough, internal versus external snapshots (snapshot-create-as, blockcommit, blockpull), live migration

serialexperimentslainnnn Skill

Use when the lifecycle of a model you train and own runs as production engineering — versioning datasets and training runs with DVC or lakeFS, tracking experiments in MLflow or Weights & Biases, promoting artifacts through a model registry with model cards, stages and approval, orchestrating training pipelines with Airflow, Kubeflow, Metaflow, Prefect or Dagster, a feature store (Feast) and train/serve skew, batch versus online versus streaming serving with shadow and canary rollout and model rollback to the previous weights and preprocessing, detecting data drift versus concept drift with Evidently when the label arrives late or never, proxy metrics and feedback loops where the model shapes its own future data, retraining triggered by schedule, threshold or event, training versus inference cost, model retirement, or fairness and bias measured as a system property.

serialexperimentslainnnn Skill

Use when authoring, reviewing or debugging Claude Code Agent Skills — SKILL.md files, frontmatter fields, description triggers, skill activation problems, catalog organization, or deciding whether something belongs in a skill, CLAUDE.md or a subagent.

serialexperimentslainnnn Skill

Provider-agnostic engineering standards for product code backed by an LLM. Use when deciding whether a task needs an LLM at all, picking a model per task by capability/cost/latency and routing by difficulty, versioning prompt templates in the repo, enforcing JSON Schema structured output instead of parsing prose, budgeting the context window and ordering it for prefix-cache hits, streaming and cancellation UX, retries/timeouts/degradation and per-user spend caps, tokens-per-request as a first-class metric with prompt-version tracing, defending against prompt injection and treating model output as untrusted input, or testing non-deterministic behaviour with golden cases and contract assertions.

ussumant Bundle

Fable Low Power mode — token-saving profile when Anthropic usage limits run low: the frontier model acts as decision maker, judge, and orchestrator ONLY; every execution step routes to cheaper models. Auto-armed by the limit-sense hook when the 5h/7d windows cross thresholds; also on demand. Use for: "low power", "save Fable", "limits are low", "running out of quota", "/fable-low-power", "conserve tokens", "battery mode".

serialexperimentslainnnn Skill

Building a language or a language processor, and deciding first whether you need one at all. Use when designing an internal (embedded) or external DSL, writing a lexer and parser by hand or with a generator (ANTLR .g4 grammars, Tree-sitter grammar.js and node-types.json, lex/flex .l and yacc/bison .y files, PEG parsers with pest .pest / peg.js / Lark, parser combinators like nom, chumsky, megaparsec or FParsec), building an AST and its source spans, name resolution and scoping, a type checker or Hindley-Milner inference, choosing a backend (LLVM IR and llvm-sys/inkwell/llvmlite, Cranelift, WebAssembly as a target, transpiling to another language, or a tree-walking versus bytecode interpreter), deciding whether a JIT is worth it, designing compiler diagnostics with spans, labels and fix-its (ariadne, codespan-reporting, miette), snapshot-testing compiler output, fuzzing a parser, building a conformance suite, or shipping the tooling a language needs to be usable - a Language Server Protocol server, formatter,

serialexperimentslainnnn Skill

Use when writing, reviewing or operating Elixir and Erlang/OTP systems on the BEAM - .ex/.exs/.erl/.hrl/.heex files, mix.exs, mix.lock, .formatter.exs, .credo.exs, rebar.config, GenServer/Supervisor/Task/Agent/Registry modules, ETS and :persistent_term, GenStage/Broadway/Flow pipelines, Ecto schemas migrations and Ecto.Multi, Phoenix contexts controllers channels and LiveView, Plug pipelines, ExUnit tests with Mox or StreamData, mix release and vm.args, distributed Erlang cookies epmd and TLS distribution, Dialyzer/Dialyxir, sobelow, or :telemetry and recon instrumentation.

serialexperimentslainnnn Skill

Umbrella skill for inherited systems - what to do with a system before touching its code, and the router to the platform skill that owns it. Use when facing a system nobody can rebuild from source, a production binary that does not match the repository, a lost or unreproducible build, code archaeology on an undocumented estate, choosing between the R strategies (rehost, replatform, refactor, rearchitect, rebuild, replace, retain, retire) for one system, a runtime or vendor that has stopped shipping updates with no upgrade path, business rules that exist only in the code and in one person about to retire, auto-translation or LLM-assisted rewrite of an old codebase, deciding to freeze and contain a system on purpose, or identifying which legacy platform skill applies to COBOL, RPG, MUMPS, VB6, Delphi, ColdFusion, Struts, Forms, ABAP or a proprietary Unix.

yuyuliuliu Skill

Use when the user pastes a long methodology/spec document (AGENTS.md, AI_USAGE.md, role-play prompt, engineering constitution, workflow spec) and wants it "added to the project" or "made into rules". This skill prevents two failure modes — (1) copying the document wholesale, which buries a near-beginner under enterprise-grade process, and (2) fabricating details the document itself forbids inventing. It lands the document as a locally-executable rule file bound to the user's real environment, archives the original verbatim, and de-duplicates overlap with existing rule files so rules live in exactly one place.

serialexperimentslainnnn Skill

The application landscape of an organization, not the design of one system. Use when building or repairing an application inventory (owner, criticality, cost, lifecycle, dependencies), running application portfolio rationalization with the Gartner TIME model (tolerate, invest, migrate, eliminate), choosing a modernization strategy from the R taxonomy (rehost, relocate, replatform, repurchase, refactor/re-architect, retire, retain), adopting or refusing TOGAF (Standard 10th Edition, ADM) and ArchiMate 3.2 and checking whether its licence lets you publish the notation, evaluating Archi, Structurizr, SAP LeanIX, Ardoq or Bizzdesign as an EA repository, building a business capability map as a stable mapping axis, setting a technology standard and the exception process with a mandatory expiry date, publishing a technology radar (Build Your Own Radar, adopt/trial/assess/caution rings), deciding build versus buy versus SaaS with exit cost and vendor lock-in, governing the integration landscape (point-to-point versus

serialexperimentslainnnn Skill

Use when a backend, runtime or system is slow and someone must prove why — defining a latency objective as percentile plus concurrency plus hardware, tail latency at p99 and p99.9, coordinated omission in load generators, open versus closed workload models, wrk2 -R, vegeta -rate, k6 constant-arrival-rate and ramping-arrival-rate executors, Gatling injectOpen versus injectClosed, JMeter Open Model Thread Group, Locust, HdrHistogram, the USE method (utilization, saturation, errors) and the RED method (rate, errors, duration), perf record and perf script, FlameGraph flamegraph.pl and stackcollapse, bpftrace and eBPF tracing, perf_event_paranoid, go tool pprof and net/http/pprof, py-spy, async-profiler and JFR, continuous profiling with Pyroscope, Parca or Grafana Profiles Drilldown, OTLP profiles and the OpenTelemetry eBPF profiler, sampling versus instrumentation overhead, CPU and allocation profiles, Amdahl's law, Little's law and queueing saturation near full utilization, latency budgets split across services

johnnylin2121 Skill

A股每日复盘工作流——涵盖盘前观察清单制定、盘中实时验证、盘后完整复盘的全流程。包括数据获取（东方财富API/cn-financial-mcp）、本地文件记录、Obsidian归档、连贯性验证、假设检验、交易计划审查。当用户要求做复盘、盘中验证、制定观察清单、审查交易计划时使用此skill。

johnnylin2121 Bundle

Amazon listing optimization workflow. Use when the user wants to analyze competitor listings to extract core keywords, write optimized titles + bullet points with keyword embedding, and generate backend search terms. Supports any Amazon marketplace. Handles browser-based crawling, keyword frequency/position analysis, and listing localization. Triggered when user mentions "亚马逊 listing", "竞品分析", "关键词", "标题五点", "后台搜索词", or similar requests.

serialexperimentslainnnn Skill

Standards for choosing and justifying an internal architecture style. Use when deciding modular monolith vs distributing, applying layered, hexagonal/ports-and-adapters, clean or onion architecture, event-driven, pipes-and-filters, plugin/microkernel or space-based styles, drawing module boundaries and dependency rules, bounded contexts and ubiquitous language, CQRS and event sourcing, ADRs (Nygard/MADR templates), C4 model diagrams and Structurizr DSL, ArchUnit or dependency-cruiser fitness functions, ISO/IEC 25010 quality attributes and quality-attribute scenarios, or diagnosing big ball of mud, anemic domain model, shared database and excessive-layering antipatterns.

icerain-cmd Skill

Use when an agent must classify a task as research, code, or hybrid, calibrate process depth, select the smallest sufficient route across the Humanities and Engineering cores, handle failed gates, and preserve a resumable workflow.

icerain-cmd Skill

Use when a researcher has an approved argument map, source-grounded notes, interpretations, and draft passages that need organization and logical repair without delegating the original scholarly contribution to the agent.

putervision Skill

Teaches the agent to use the Behavior MCP server for ~60Hz in-browser behavior trees, triggers, and recordings.

putervision Skill

Teaches the agent to process, ingest, analyze, and compare WebM, MP4, and GIF video recordings using vision-memory-mcp and state-memory-mcp.

putervision Skill

Teaches the agent to use the WebCrypt MCP server for AES-256-GCM symmetric encryption, RSA-4096 hybrid encryption, key generation, digital signatures, hashing, and post-quantum cryptography.

putervision Skill

Teaches the agent to use the Spatial World Model MCP server to track entities, 3D/2D positions, spatial relationships, object permanence, and movement simulation.

putervision Skill

Teaches the agent to use the state-memory-mcp MCP server to track workflow state, tasks, decisions, blockers, artifacts, plans, milestones, and their semantic relationships in a persistent graph database.