Context & AI engineering
25,729 rows, by stacks then stars
1,865 of these skill files have been read, by 1,328 distinct authors, and what they tell an agent counted. What Context & AI engineering authors agree on, and what they forbid
kjuhwa/skills-hub/skills/devops/worktree-env-file-isolation Skill
0★ repoPer-git-worktree environment files (.env.worktree) with deterministically derived unique database names and ports, sharing one Postgres container.
kjuhwa/skills-hub/skills/electron/electron-bun-dev-orchestrator Skill
0★ repoCross-platform Electron dev launcher that parallel-builds main/preload/MCP subprocess bundles, waits for file stability, watches with esbuild's JS API, and spawns Vite + Electron - replaces a fragile tree of npm scripts.
Karpathy llm coding guidelines
kjuhwa/skills-hub/skills/engineering/karpathy-llm-coding-guidelines Skill
0★ repoUmbrella behavioral guideline that bundles the four Karpathy principles — Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution — to reduce common LLM coding mistakes. Use when you want all four applied at once.
kjuhwa/skills-hub/skills/engineering/setup-engine Skill
0★ repoConfigure the project's game engine and version. Pins the engine in CLAUDE.md, detects knowledge gaps, and populates engine reference docs via WebSearch when the version is beyond the LLM's training data.
kjuhwa/skills-hub/skills/fastapi-streaming-endpoints/fastapi-sse-mcp-protocol Skill
0★ repoHost a hosted MCP (Model Context Protocol) server over FastAPI's StreamingResponse using API-key auth, a session dict, and a tool registry surfaced via the 2025-03-26 Streamable HTTP spec.
kjuhwa/skills-hub/skills/flutter/svg-image-asset-embedding Skill
0★ repoFlutter SVG asset handling with flutter_svg in desktop UI.
Git worktree isolation per thread
kjuhwa/skills-hub/skills/git/git-worktree-isolation-per-thread Skill
0★ repoUse git worktrees to isolate filesystem state per conversation thread, avoiding conflicts between concurrent edits
Tensor memory accelerator descriptor construction
kjuhwa/skills-hub/skills/gpu-kernels/tensor-memory-accelerator-descriptor-construction Skill
0★ repoBuild Hopper/Blackwell TMA descriptors at host time via cuTensorMapEncodeTiled so kernels can issue async tiled loads without per-address arithmetic.
kjuhwa/skills-hub/skills/inference/lm-prompt-kv-cache-reuse Skill
0★ repoPre-compute and cache LM KV state for a fixed prompt, then reuse across multiple target generations
kjuhwa/skills-hub/skills/input/rdev-global-input-hook Skill
0★ repordev crate for global keyboard/mouse event hooking and input simulation.
Claude cli grace kill after success result
kjuhwa/skills-hub/skills/integration/claude-cli-grace-kill-after-success-result Skill
0★ repoWhen driving `claude -p --output-format stream-json`, treat a `result:success` event as completion even if the child doesn't self-exit; arm a 15s grace timer to kill, and return ok=true regardless of exit code.
kjuhwa/skills-hub/skills/integration/computer-use-agent Skill
0★ repoBuild an agent that controls a computer (screenshots, clicks, keyboard) using ComputerTool.
Mcp multi transport config pydantic
kjuhwa/skills-hub/skills/integration/mcp-multi-transport-config-pydantic Skill
0★ repoSingle Pydantic config that supports three MCP transports (stdio, sse, streamable-http) with per-transport required-field validation in a model_validator(after) so users get a clear error when the wrong combination is set.
Mcp streamable http detached httpx client
kjuhwa/skills-hub/skills/integration/mcp-streamable-http-detached-httpx-client Skill
0★ repoInject a caller-owned httpx.AsyncClient into the official MCP streamable_http transport via httpx_client_factory, using a no-op async-context-manager so the transport never closes the underlying client on exit.
kjuhwa/skills-hub/skills/llm-agents/agent-handoff-routing Skill
0★ repoRoute a conversation to a specialist agent via handoff, making that agent own the next turn.
kjuhwa/skills-hub/skills/llm-agents/chain-of-thought-prompting Skill
0★ repoApply chain-of-thought prompting (zero-shot CoT, few-shot CoT, Self-Consistency) to improve LLM multi-step reasoning.
kjuhwa/skills-hub/skills/llm-agents/deterministic-agent-chain Skill
0★ repoChain multiple agents sequentially in code, feeding each agent's output as the next agent's input.
kjuhwa/skills-hub/skills/llm-agents/dispatching-parallel-agents Skill
0★ repoUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies
Gui agent sft qwen2vl os kairos
kjuhwa/skills-hub/skills/llm-agents/gui-agent-sft-qwen2vl-os-kairos Skill
0★ repoSFT Qwen2-VL-7B on the OS-Kairos GUI dataset using LLaMA-Factory, converting data to ShareGPT format, and running inference.
kjuhwa/skills-hub/skills/llm-agents/handoff-input-filter Skill
0★ repoFilter or transform the conversation history passed to a specialist agent on handoff.
kjuhwa/skills-hub/skills/llm-agents/handoff-with-typed-input Skill
0★ repoProvide structured data from the LLM to an on_handoff callback via input_type on handoff().
kjuhwa/skills-hub/skills/llm-agents/hermes-acp-agent-server Skill
0★ repoExpose an LLM agent via the Agent Client Protocol (VS Code / Zed / JetBrains) with per-session state and MCP injection.
Hermes anthropic prompt cache system and 3
kjuhwa/skills-hub/skills/llm-agents/hermes-anthropic-prompt-cache-system-and-3 Skill
0★ repoPlace Anthropic 4 cache_control breakpoints as system + last 3 messages so multi-turn chats keep a hot cache.
Hermes context compression handoff
kjuhwa/skills-hub/skills/llm-agents/hermes-context-compression-handoff Skill
0★ repoCompress long agent conversations with a handoff-framed summary that prevents re-answering resolved turns.
kjuhwa/skills-hub/skills/llm-agents/hermes-cron-agent-scheduler Skill
0★ repoRun an LLM agent on a cron schedule with inactivity timeout, silent marker, and cross-platform delivery.
kjuhwa/skills-hub/skills/llm-agents/hermes-delegate-subagent Skill
0★ repoSpawn child LLM agents with restricted toolsets, isolated context, and summary-only parent feedback.
Hidden state capture layer hook
kjuhwa/skills-hub/skills/llm-agents/hidden-state-capture-layer-hook Skill
0★ repoCapture hidden states from specified transformer layers by wrapping them with a small proxy that records the output and forwards all other attribute access to the original layer.
Llm api zero few shot prompting
kjuhwa/skills-hub/skills/llm-agents/llm-api-zero-few-shot-prompting Skill
0★ repoCall Qwen (DashScope) or OpenAI APIs for zero-shot and few-shot prompting, including streaming responses.
kjuhwa/skills-hub/skills/llm-agents/llm-as-judge-loop Skill
0★ repoRun an agent in a loop with an evaluator agent that provides feedback until the output meets a quality threshold.
kjuhwa/skills-hub/skills/llm-agents/output-guardrail-with-agent Skill
0★ repoUse a fast/cheap agent to validate the main agent's output before returning it to the user.
kjuhwa/skills-hub/skills/llm-agents/parallel-agent-runs Skill
0★ repoRun multiple agent instances concurrently with asyncio.gather() and pick the best result.
Plan based multi step task decomposition
kjuhwa/skills-hub/skills/llm-agents/plan-based-multi-step-task-decomposition Skill
0★ repoPlan mode for multi-step agent tasks — exploration via subagent, gated planning with user confirmation, strict execution loop, and adversarial verification.
kjuhwa/skills-hub/skills/llm-agents/subagent-driven-development Skill
0★ repoUse when executing implementation plans with independent tasks in the current session
Macos privilege scripts embedding
kjuhwa/skills-hub/skills/macos/macos-privilege-scripts-embedding Skill
0★ repoEmbed shell scripts via include_dir!() for privilege elevation + system operations.
kjuhwa/skills-hub/skills/mcp-integration/hosted-mcp-tool Skill
0★ repoLet OpenAI's Responses API call a remote MCP server on your behalf using HostedMCPTool.
Mcp pool http proxy for subprocess sdks
kjuhwa/skills-hub/skills/mcp-integration/mcp-pool-http-proxy-for-subprocess-sdks Skill
0★ repoExpose a pool of MCP sources (Linear, GitHub, Notion, ...) to Codex/Copilot SDK subprocesses through a single Streamable-HTTP MCP endpoint, instead of having each subprocess open independent connections.
kjuhwa/skills-hub/skills/mcp-integration/mcp-sse-streamable-http Skill
0★ repoConnect an agent to a remote MCP server via HTTP+SSE or Streamable HTTP transport.
kjuhwa/skills-hub/skills/mcp-integration/mcp-stdio-server Skill
0★ repoConnect an agent to a local MCP server running as a subprocess over stdio.
Mcp tool intent metadata injection
kjuhwa/skills-hub/skills/mcp-integration/mcp-tool-intent-metadata-injection Skill
0★ repoInject a hidden _intent / _displayName field into every MCP tool schema via a fetch interceptor so large-response summarization and UI rendering have context that the LLM never sees.
Session scoped mcp stdio server
kjuhwa/skills-hub/skills/mcp-integration/session-scoped-mcp-stdio-server Skill
0★ repoShip a per-session MCP stdio server bundled as CJS that provides session-scoped tools (SubmitPlan, config_validate) and uses stderr __CALLBACK__ lines to pause execution for async host-process callbacks.
Fastmcp stdio http dual transport
kjuhwa/skills-hub/skills/mcp/fastmcp-stdio-http-dual-transport Skill
0★ repoSingle MCP server entry point that supports both STDIO (default) and Streamable HTTP / SSE transports via a --http flag, using FastMCP + Starlette + uvicorn, with localhost-binding-by-default and a warning banner on non-localhost binds.
kjuhwa/skills-hub/skills/memory-management/layered-memory-tiering-system Skill
0★ repoHierarchical L1/L2/L3/L4 memory architecture for long-running LLM agents, anchored by action-verified facts and explicit garbage-collection rules.
kjuhwa/skills-hub/skills/misc/write-a-skill Skill
0★ repoCreate new agent skills with proper structure, progressive disclosure, and bundled resources. Use when user wants to create, write, or build a new skill.
Sequential two phase training orchestrator
kjuhwa/skills-hub/skills/ml-ops/sequential-two-phase-training-orchestrator Skill
0★ repoDrive a two-phase training pipeline (e.g. tokenizer then predictor) with one command, CLI skip flags, and auto-skip if an earlier phase's best_model already exists.
Tts trailing silence and hallucination trim
kjuhwa/skills-hub/skills/ml-ops/tts-trailing-silence-and-hallucination-trim Skill
0★ repoPost-process TTS output to cut long internal silences plus any trailing hallucinated noise with a cosine fade-out.
Mega moe fusion with symmetric memory
kjuhwa/skills-hub/skills/moe-optimization/mega-moe-fusion-with-symmetric-memory Skill
0★ repoFuse EP dispatch, FP8×FP4 GEMM×2, SwiGLU, and EP combine into a single kernel that reads across ranks via symmetric-memory (same VA on every rank) to overlap NVLink collectives with compute.
Fhs wrapper mcp server support
kjuhwa/skills-hub/skills/nix/fhs-wrapper-mcp-server-support Skill
0★ repoWrap a NixOS application derivation in buildFHSEnv to provide a FHS-compliant filesystem view. Enables subprocess execution of dynamically linked native binaries (like MCP servers) that expect /usr, /lib, /bin at standard paths.
kjuhwa/skills-hub/skills/observability/agent-tracing-spans Skill
0★ repoWrap multi-agent workflows in named traces and custom spans for debugging in the OpenAI Traces dashboard.
Threadpool benchmark with warmup
kjuhwa/skills-hub/skills/observability/threadpool-benchmark-with-warmup Skill
0★ repoRun a concurrent HTTP load test against an LLM server with a warmup phase, tqdm progress on completion, and per-prompt result aggregation — all using the stdlib ThreadPoolExecutor.
kjuhwa/skills-hub/skills/operations/claude-cli-unattended-wrapper Skill
0★ repoSpawn `claude -p` non-interactively from a Node parent process so a long-running loop can drive slash commands without any human at the keyboard. Layers a prompt tempfile, a piped "y\n" stream, hook-bypass env vars, and a stream-json consumer to keep the child from ever blocking.
Multi format native package orchestrator
kjuhwa/skills-hub/skills/packaging/multi-format-native-package-orchestrator Skill
0★ repoOrchestrate building a single Electron app source into .deb, .rpm, AppImage, and Nix outputs from one script. Handles architecture detection, distro-family detection, format-specific build dispatch, and shared staging.
kjuhwa/skills-hub/skills/safety/input-guardrail-agent Skill
0★ repoBlock off-topic or policy-violating user input before the main agent runs using @input_guardrail.
R judge agent safety evaluation
kjuhwa/skills-hub/skills/safety/r-judge-agent-safety-evaluation Skill
0★ repoUse R-Judge to evaluate an LLM agent risk awareness over multi-turn ReAct records, producing a risk description and safe/unsafe label.
kjuhwa/skills-hub/skills/security/mcp-subprocess-env-filtering Skill
0★ repoWhen spawning a stdio MCP server (or any untrusted subprocess), allowlist/denylist environment variables so app-level auth tokens and third-party API keys don't leak into the child process.
kjuhwa/skills-hub/skills/testing/benchmark Skill
0★ repoPerformance regression detection using the browse daemon. Establishes baselines for page load times, Core Web Vitals, and resource sizes. Compares before/after on every PR. Tracks performance trends over time. Use when: "performance", "benchmark", "page speed", "lighthouse", "web vitals", "bundle size", "load time". (gstack) Voice triggers (speech-to-text aliases): "speed test", "check performance".
Llm integration test failure diagnosis
kjuhwa/skills-hub/skills/testing/llm-integration-test-failure-diagnosis Skill
0★ repoWire an LLM-based diagnosis step into your CI for integration-test failures — summarize noisy logs, identify probable root cause, post findings to the code-review/PR UI. Based on Google's Auto-Diagnose system (90%+ accuracy on 71-case study, adopted across 52k tests).
Credential prompt dialog abstraction
kjuhwa/skills-hub/skills/ui/credential-prompt-dialog-abstraction Skill
0★ repoCredential dialog abstraction across platforms via ui_interface::msgbox callback.
kjuhwa/skills-hub/skills/workflow/agent-feedback-loop Skill
0★ repoRun an agent repeatedly in a while loop with user or evaluator feedback until a quality threshold is met.
Async subagent resume on missing artifact
kjuhwa/skills-hub/skills/workflow/async-subagent-resume-on-missing-artifact Skill
0★ repoWhen an async subagent is reported "completed" but its final message hints at unfinished work (e.g., "Now let me build…"), verify the expected on-disk artifact before moving on; if missing, use SendMessage to the agent id to finalize rather than re-dispatching a fresh agent.
Delegate bulk scans to explore subagent
kjuhwa/skills-hub/skills/workflow/delegate-bulk-scans-to-explore-subagent Skill
0★ repoScan-heavy slash commands (repo imports, full-project extracts, corpus-wide refactors) must delegate bulk file/web scanning to an Agent(subagent_type=Explore) in a single call, then synthesize drafts in the main thread. Prevents the 50-70 tool call thrashing that fragments conversation history.