GITHUB AGENT SKILLS / 技能收录

给你的 AI 助手
装上专业技能

我们从 GitHub 持续收录、审核并分类优质的 Agent Skills,按职业与使用场景整理,复制一行命令即可安装到 Claude Code、Cursor 等智能体。

01 /

按职业浏览 Agent Skills

为不同岗位挑选的高频技能

全部职业
02 /

按使用场景浏览 Agent Skills

从你要完成的任务出发

全部场景
03 /

按创作者浏览 Agent Skills

按 GitHub Stars 排序的优秀创作者

全部创作者

全部 Skills 482556

nubjs
nubjs/nub

research-thread

Use when the deliverable is a FACT — not code, not a design, not a gap catalog. A prior-art survey, "how has X already been solved", "what did they try and abandon", a long empirical sweep, or any non-trivial question whose answer will be cited for months. Produces a living write-up at `wiki/research/<topic>.md`. Auto-triggers on "research this", "prior art", "survey", "find out how X works", "what's the state of the art", "dig into whether".

★ 4.4k
nubjs
nubjs/nub

release

Cut a Nub patch release end-to-end in one invocation. Invoke (via the Skill tool) once a release thread's targeted fixes are ALL landed on `main` and CI-green. Encodes the full runbook: pick the version (patch bump in the 0.0.x/0.1.x pre-release regime), audit `@nubjs/types`, run `make version` + `make version-check`, commit + push to `main`, write the FACTUAL + NEUTRAL release notes from the full changeset, then dispatch release.yml with publish=true and the notes as its input (that dispatch runs the 8-platform build → glibc and pre-publish native gates → immutable 32-asset DRAFT release carrying the notes → signed packslip → npm OIDC STAGE → a held run until the maintainer approves with 2FA → the `v<ver>` tag and `v0` pushed with the release deploy key → stable GitHub Release), `stager await` the approval with no timeout, then comment the version + release link on every closed issue + merged PR the release ships (mandatory maintainer hygiene). Do NOT cut until all fixes are green.

★ 4.4k
nubjs
nubjs/nub

probe-platforms

Bring a test harness or probe up on a NEW operating system (Windows or Linux), or debug one that fails there. Invoke before the first run on a platform, and whenever a harness works on one OS but not another. Carries the bring-up ladder that avoids spending hours learning single booleans, the Windows spawn/path/disk faults that each present as something else entirely, the Linux Landlock and node-layout traps, and the remote-shell mechanics (PowerShell-over-SSH quoting, pkill matching your own command) that waste a loop each time they are rediscovered.

★ 4.4k
nubjs
nubjs/nub

pm-perf-tracing

Performance-trace Nub package-manager installs using the existing phase timings, structured diagnostics, and sampling-profiler workflow. Use when an install or package-manager operation is unexpectedly slow and the bottleneck needs to be localized before changing code.

★ 4.4k
nubjs
nubjs/nub

plan-thread

Use when an effort's APPROACH isn't settled yet and needs collaborative design BEFORE implementation — fleshing out how a feature/change should work, weighing options, resolving open questions with the human. The design-in-progress thread profile. Auto-triggers on "spin up a plan", "let's design", "plan thread", "how should we approach", "think through the approach".

★ 4.4k
nubjs
nubjs/nub

orchestrator

暂无简介

★ 4.4k
nubjs
nubjs/nub

nub-charts

Build a performance chart for nubjs.com — the SVG bar figures in blog posts, docs pages and social posts (a runtime augmentation against plain node, an install or dispatch comparison, a cross-tool ranking). Use whenever asked for a graphic, chart, graph or figure showing a benchmark result, whenever adding a measured claim to a docs or blog page that deserves a picture, and whenever regenerating an existing figure with fresh numbers. Covers the house visual system (720px, the light/dark/opaque triple, the Nub palette and type), the three chart forms and which one fits, the Figure wiring and the alt/caption conventions, the benchmark provenance the numbers must have, and the layout traps that only show up once the SVG is rasterized.

★ 4.4k
nubjs
nubjs/nub

linux-vm-test

Run ad-hoc Nub tests and debugging probes on real local Linux guests. Use for Linux-specific sandbox, kernel, distro, architecture, package, or runtime behavior that Docker cannot reproduce faithfully. Selects persistent Lima for source-mounted iteration or ephemeral Tart for clean-image reproduction.

★ 4.4k
nubjs
nubjs/nub

lat-md

Search, read and maintain nub's knowledge graph under wiki/ with the `lat` CLI — the design and research corpus, cross-linked and checked. Invoke (via the Skill tool) before designing or changing anything non-trivial, to find the decision that already governs it instead of re-deriving it; and after any change that alters architecture, behavior or test coverage, because `lat check` is a CI gate and a stale wiki link now fails the build. Carries the commands, the section-id syntax, the two rules that make a section valid, and the three nub-specific traps — never run bare `lat init`, never create `.agents/skills/`, and Rust symbols inside a `mod` block cannot be linked.

★ 4.4k
nubjs
nubjs/nub

implementation-thread

Take a single task or issue through the COMPLETE engineering workup — grill the brief with the human → plan → review the plan → implement → review the implementation (multi-lens where the blast radius warrants) → open a PR → await CI → fix CI → check for and integrate external reviews → RETURN CONTROL before merge. Invoke (via the Skill tool) whenever the user says "implementation-thread", or asks to drive a task/issue all the way through to a held PR. The defining shape: the effort is owned end-to-end by ONE sub-agent (an L1) that spawns its OWN L2 sub-agents for design and multi-lens review, carries continuity across phases, and PAUSES — comes to rest to surface a question UP — whenever a decision the human owns arises. It opens by loading the `grilling` skill and running its round-based INTERVIEW whenever the change writes substantial new code, so the human settles the design tree before a line of it exists. It returns control to the orchestrator BEFORE merging unless explicitly told "merge it" / "all the way through". It governs how ONE implementation effort runs internally, so it composes with — rather than replaces — whatever is driving the wider campaign.

★ 4.4k
nubjs
nubjs/nub

gcloud-vm

暂无简介

★ 4.4k
nubjs
nubjs/nub

download-stats

Generate download-stats CSVs and a chart for nub across its distribution channels (npm + GitHub release assets, which subsume Homebrew). Invoke (via the Skill tool) when asked to check, chart, or export download numbers, install counts, or adoption stats. Encodes the channel map, which numbers may be added together, each channel's real granularity and lag, and the checksum-asset inflation gotcha.

★ 4.4k
nubjs
nubjs/nub

disk-reduction

Reclaim disk on the maintainer's Mac when the volume is full or filling — ENOSPC, "no space left on device", a failed build or agent harness, or a routine sweep of Rust build residue. Invoke whenever free space is the problem. Covers the two families that actually hold the space (worktree-private `target/` dirs and orphaned content-hashed shared-target buckets under `~/.cache/nub`), the safety gates that keep a live bucket and the installed `nub-dev` binary alive, and why `du` and `df` disagree by tens of GB on APFS. For CPU load, orphaned processes, or a build hung on a target lock, use `cpu-reduction`; for not creating the residue, `rust-build-hygiene`.

★ 4.4k
nubjs
nubjs/nub

dev-loop

Build and test nub during development. Invoke (via the Skill tool) whenever you need to compile the dev `nub` binary, set up a worktree for fast incremental iteration, run a specific test file or a single test, or get oriented in the codebase (the crate map). Encodes the measured fast-build loop: the `fast` profile built through `scripts/rust-build.sh`, which shares ONE CARGO_TARGET_DIR across worktrees (`~/.cache/nub/shared-target`) so deps are reused and only the workspace crates recompile — but auto-isolates a worktree to a private target dir the moment it diverges a depended-on crate (vendor/aube, nub-core, …), which is when a shared dir would clobber a sibling and fail with a phantom compile error on correct source (the `rust-build` skill). A shared cross-worktree compiler-WRAPPER cache (sccache) was separately measured to give 0% Rust speedup and is NOT used. Covers the real incantations (`cargo build -p nub-cli --profile fast`, `make install-dev`, `make addon-fast`), the test invocations, and the exact CI cheap gates. SCOPE — the local loop is the EXCEPTION, not the default: a cold build, `clippy --all-targets --all-features`, a full `cargo test`, a `release` build or a multi-fixture sweep goes OFF-HOST, via a CI-run test or the `remote-build` skill. This skill covers only what must stay local — the ~5s warm rebuild you are actively iterating against, and macOS-native checks.

★ 4.4k
nubjs
nubjs/nub

cpu-reduction

Diagnose and clear CPU, memory, and disk contention on the maintainer's dev host. Invoke when the machine is slow, load or swap is high, disk is filling, a Rust build hangs on a target lock, a benchmark needs a quiet host, stale worktree targets have accumulated, or orphaned build/test processes are suspected. Covers safe worktree-target cleanup, detached Rust builds, orphaned Node tests and synthetic load generators, the preserve list, and the measure-to-verify loop. For preventing orphan builds, use `rust-build-hygiene`.

★ 4.4k
nubjs
nubjs/nub

corpus-sweep

暂无简介

★ 4.4k
nubjs
nubjs/nub

corpus-remeasure

暂无简介

★ 4.4k
nubjs
nubjs/nub

ci-watch

Watch GitHub Actions CI correctly with the gh CLI — block until a run / PR check rollup is TRULY terminal, then trust the exit code. Invoke (via the Skill tool) whenever you need to wait on CI after REQUESTING a pull-request run (`gh pr edit <n> --add-label ci` — PR CI is opt-in, so a push starts nothing), after a push to main, or after a tag, and act on the result (merge-on-green, release-on-green, fail-fast on red). Encodes the premature-exit pitfall (raw `gh run watch` / `gh pr checks --watch` exit 0 while the run is still QUEUED with no jobs registered, and exit non-zero on a transient API blip) and the blessed fix: `scripts/ci-watch.ts`, which waits for the target to EXIST, polls authoritative terminal status, fails fast on the first failing check, and exits with a status the orchestrator can trust. Run it as a detached run_in_background task.

★ 4.4k