★ Curated agent skills ★
A curated collection of genuinely useful agent skills — one representative per capability, vetted for quality, categorized by function.
85 skills · 7 categories · 8 ⭐ specials · 6 link-only
View on GitHubClarifies what the user actually wants, one question at a time, compares 2-3 approaches, and gets a design approved before anything is built.
Stops build-first-guess-later work on new features and projects; it is the process that designed this very repo.
Terse caveman voice for agent replies: answer first, fluff gone, every technical fact kept, and it stays on until told to stop.
Cuts output-token waste on every reply without losing substance; the communication-side counterpart to ponytail's lean code.
Compresses memory files such as CLAUDE.md or todo lists into caveman format to save input tokens, keeping a readable backup outside the tree.
Long-lived context files are re-read every session; compressing them is a permanent input-token saving rather than a one-off.
Builds and sharpens a project's domain model: maintains a living glossary of codebase terminology and records architecture decisions as ADRs.
Shared vocabulary keeps agents and humans from talking past each other; this makes terminology a maintained artifact, not folklore.
Finds and installs agent skills from the open ecosystem through the Skills CLI and the skills.sh registry.
Skill discovery is a different job from skill authoring — and the front door for growing this very collection.
A relentless one-question-at-a-time interview that stress-tests a plan or design until every branch of the decision tree is resolved.
The sharp end of requirements work: it surfaces the decisions brainstorming leaves unmade, before they turn into rework.
The reusable interview engine behind grill-me: grills the user relentlessly about a plan, decision, or idea, one question at a time.
Grill-me is the session wrapper; this is the portable technique that other workflows can invoke on any decision.
Survey any codebase as a senior advisor and produce prioritized, self-contained implementation plans for other models/agents to execute — read-only on source code, never implements anything itself.
The audit-first advisor loop: recon, then a 9-category parallel audit, findings vetted and prioritized by leverage, and plans written for the weakest plausible executor, executed in an isolated worktree — distinct from writing-plans (plans from a given spec) and subagent-driven-development (executes an existing plan).
Anthropic's playbook for designing and building high-quality MCP servers, with reference guides and an evaluation harness.
The owner builds connectors and MCP-backed tools; this is the reference guide from the source.
Lazy senior-dev mode: the smallest change that fully solves the task, explained in one read.
Cuts over-engineering, tokens, and cost on everyday coding without dropping safety or tests.
One-shot whole-repo audit for bugs, security holes, slow paths, untested risky code, and deletable bloat.
A fast health check for inherited or fast-grown codebases, ranked by impact.
Lists every deliberate shortcut: marker in the code as a debt ledger.
Keeps intentional shortcuts visible so they get repaid instead of forgotten.
Quality review of a change: correctness, safety, load behavior, tests, and whether every line earns its place.
Catches risky or bloated diffs before they ship, explained in plain language.
Anthropic's guide and tooling for creating, evaluating, and iterating on agent skills, with eval and description-tuning scripts.
This repo curates skills for a living; the official skill-authoring workflow belongs in it.
Suggests manual context compaction at logical task-phase boundaries, so long sessions compact deliberately instead of at arbitrary auto-compaction points.
Long agent sessions die at context limits; compacting at phase boundaries keeps the working thread coherent across a whole build.
Executes an implementation plan with fresh subagents per task, with spec and quality review gates between tasks.
The execution engine for writing-plans in multi-agent harnesses; how the owner's bigger builds actually get run.
A root-cause-first debugging method: reproduce, isolate, and prove the cause before changing any code.
Ends guess-and-patch debugging loops; pairs with this project's verify-against-real-behavior rule.
Guides red-green-refactor TDD: write the failing test first, make it pass with the smallest change, then refactor.
Proven discipline from the most-starred source in this collection; keeps agent-written code verified by construction.
Three.js scene setup done right: scenes and fog, cameras, the renderer and tone mapping, the animation loop, object hierarchies, transforms, core math, and cleanup, on WebGL or WebGPU/TSL.
The fundamentals entry of the Three.js set: scene, camera, and renderer mistakes are where 3D web work goes wrong first.
Sets up isolated git worktrees so parallel features, reviews, and experiments never collide in one working tree.
The standard companion to plan-driven work; makes parallel agent work on the owner's repos safe.
Vercel Engineering's React and Next.js performance rules: 70 prioritized guidelines across data fetching, bundles, rendering, and re-renders.
The canonical React performance rulebook; applies to the owner's Next.js work (Eggino) and any React build.
Requires fresh evidence (test runs, command output, live checks) before any claim that work is done or fixed.
Encodes the rule that shipped means verified; cheap insurance against false 'done' reports.
Tests local web apps with Playwright: manages the dev server, drives a real browser, and inspects console, DOM, and screenshots.
Gives agents a real-browser check for the owner's web apps and artifacts; catches what unit tests cannot see.
Cloudflare's official best practices for Workers: configuration, runtime patterns, and platform API guidance.
Nearly all of the owner's production code runs on Workers (IVAI, TradeAgent, Proxiv3); this is the vendor's own guidance.
Turns an approved spec or design into a detailed, step-by-step implementation plan that can be executed without guessing.
The missing half of the brainstorming workflow already in this repo; makes multi-step builds land first-try instead of drifting.
Creates generative art with p5.js: seeded algorithms, flow fields, and particle systems delivered as interactive HTML pieces.
The distinctive one in Anthropic's set: real generative-art capability, not another text workflow; useful for visuals and covers.
Builds crafted, interactive isometric SVG figures that explain a technical idea as a physical machine, with worked examples and a geometry-audit kit.
Precise technical illustration as explanation — a capability neither the generative-art skill nor the UI set covers.
Designs banners for social, ads, website heroes, and print, with multiple art-direction options.
Banners and covers are a recurring need for channels and launches.
Drives Blender headless or via MCP to build, modify, render, and optimize 3D models, then publishes them as fast Three.js/Next.js web experiences with modern color, shader, and compression pipelines.
The only Blender-to-web pipeline in the collection: real 3D asset craft with optimization and deploy covered, plus a runnable demo to start from.
Brand voice, visual identity, messaging frameworks, and consistency checks for branded content.
Keeps channel and product visuals and tone recognizably on-brand.
Measures color contrast with WCAG 2.x ratios and APCA, audits a codebase or stylesheet for failing pairs (including dark mode), and repairs failures with the smallest possible color change.
The executable half of color accessibility: it audits real stylesheets, computes real ratios, and fixes failures — measurement work no resident skill does.
Deep color expertise: color spaces (OKLCH, CIELAB), palette and ramp generation, harmony systems, color naming, pigment and print-vs-screen behavior, and contrast verification.
The palette data elsewhere is lookup-only; this is the craft of making and judging color — perceptually uniform ramps, naming, and verified contrast.
Design-token architecture (primitive to semantic to component), component specs, and brand-compliant slides.
Keeps multi-screen products visually consistent as they grow.
Produces 45 diagram types — architecture, sequence, ER/data models, flowcharts, Wardley maps, Sankey, Gantt and more — as self-contained HTML+SVG, with a style-guide gate that pulls theme tokens from a site URL or design system so diagrams match the brand.
The collection's only diagramming skill: no other entry makes technical or editorial diagrams, and its dependency-free HTML+SVG output is themed from the target's own tokens instead of generic AI styling.
Emil Kowalski's design-engineering philosophy as a skill: whether an element should animate at all, custom easing curves, durations, interruptible transitions, and the invisible polish details that make interfaces feel right.
Fills the collection's animation/motion-craft gap: no other entry covers animation decision-making or easing craft. It is the flagship of the source repo's 18 skills; the siblings are deliberately not imported (one per cluster), so the source repo is the pointer to the rest.
Generates new icons in one fixed, consistent house style: a rulebook-driven design engine with a headless visual self-check and a static SVG exporter, companion to the Kubai icon library (2,600+ icons for React, Vue and plain SVG).
Icon generation in a single consistent style is a job no other design entry does — ui-ux-pro-max's data covers icon selection, not creation. Listed link-only: the skill's own license is source-available, bars translation and modified redistribution, and its text is Chinese, so it cannot be vendored into this English collection.
Creates strategic HTML presentations with charts, design tokens, and responsive layouts.
Turns briefs and reports into presentable decks without design software.
Builds accessible UIs with shadcn/ui, Radix, and Tailwind, plus canvas-based visual designs.
The practical implementation partner to the design-system rules.
Design intelligence for web, mobile, and desktop UIs, with searchable local data on styles, palettes, font pairings, UX guidelines, and stacks.
Gives interface work a real design database instead of guesswork; the flagship of the design set.
Builds and leverages online communities - Discord, Slack, forums - for product growth and loyalty.
Community-led growth fits bot and channel products especially well.
Researches competitors from their URLs and produces structured profile dossiers.
Know exactly what rivals offer and how they position before you ship.
Plans what content to create: pillars, topic clusters, editorial calendars, and roadmaps.
Deciding what to publish beats publishing at random, for blogs and channels alike.
Edits and refreshes existing marketing copy: tightening, de-slopping, and sharpening without a rewrite.
Most real work is improving copy that already exists, not starting from zero.
Writes and improves marketing copy for pages: headlines, CTAs, value propositions, and hero sections.
Sharp copy is the cheapest conversion lever for landing pages and product posts.
Runs and synthesizes customer research: interviews, surveys, review mining, voice-of-customer, and JTBD.
Decisions grounded in real customer language outperform opinion.
Plans product launches and feature announcements: checklists, go-to-market, waitlists, and release strategy.
Launches are one-shot events; a checklist prevents unforced errors.
A bank of proven marketing and growth tactics for software products, matched to the situation.
The fastest way out of 'we need growth but don't know what to try'.
Applies behavioral science and mental models - anchoring, social proof, scarcity, framing - to marketing.
Explains why people buy, so offers and copy stop being guesswork.
Designs the offer itself: value framing, bonus stacks, guarantees, scarcity, and payment structure.
A better offer beats better ads; core skill for anything sold or promoted.
Creates the product-marketing context doc - positioning, audience, ICP, brand voice - that the other marketing skills build on.
The foundation of the marketing set: write the context once, reuse it in every campaign.
Creates, schedules, and optimizes social media content across the major platforms.
Directly useful for running and growing social and channel presences.
Directs film and AI video like a working director: script breakdowns, beat sheets, blocking and staging, shot lists, storyboard and motion prompts, continuity bibles, and per-model adapters.
The filmmaking-craft representative: executable direction for a camera crew or a generation queue, beyond the video skill's strategy-level coverage.
A complete local FFmpeg editing system: 42 scripted tools to cut, join, reframe, caption, remove silence, check loudness and export audio/video.
Deterministic media-file work, fully local and scriptable — nothing else in Prime Skill touches actual audio or video files.
Makes complete short films entirely in code: describe a story, pick one of 66 styles (each shipped with a demo film), and the agent directs and renders it with Node/canvas/Three.js and ffmpeg — no video generator involved.
End-to-end code-film production, a genuinely distinct job from cinematic-director's planning layer; listed link-only to keep the collection's motion cluster to one vendored representative.
Operates OpenWhispr from the terminal via the `openwhispr` CLI: manage notes, folders, transcriptions and audio, and script dictation workflows against the local desktop app or the cloud API.
Local, offline-capable dictation as an agent-operable CLI — a speech-to-text workflow nothing else in the collection covers.
Reverse-engineers how a video was made, shot by shot: classifies each shot (live action, AI-generated, 3D, 2D/stop-motion, code render, screen capture and more), detects AI edits and overlay packaging, breaks down the audio, and produces an executable remake plan.
REA's sibling for video: the only entry that analyzes and deconstructs existing videos instead of creating new ones — nothing else in the collection reads a finished video's production method.
From Jenesyx/arta-edit-studio: turns Farsi-first talking-head videos into animated Instagram Reels — this representative skill is the full-screen story format with B-roll and stickers, rendered with HyperFrames from a Whisper transcript, with no burned-in subtitles. The repo also ships the reel-setup and reel-split skills.
A Farsi-first Reels editing workflow, unique in the collection and close to the curator's own content work; link-only while the young repo matures.
Remotion's official best practices for programmatic video in React — markup, animation, captions, maps and interactivity done right.
The motion cluster's one representative: official, actively maintained engine craft, deeper than the video skill's strategy-level coverage; link-only, no license declared.
Guides AI and programmatic video production: avatars, demos, explainers, and generation pipelines.
Video is the highest-reach format; this maps the current tool landscape into workflows.
Drives a locally running VoiceStudio backend (REST API or MCP): text-to-speech with saved or designed voices, voice cloning, transcription, and automated dubbing or narration workflows.
The collection's only voice skill: local-first speech generation and cloning, no cloud TTS account — nothing else in Prime Skill makes or processes spoken audio.
Analyzes active ads from Meta, Google, and LinkedIn libraries: hooks, offers, CTAs, and claims.
Competitors publish their best-tested messaging in ad libraries for free.
Free, local, read-only access to platform content: search and read posts, comments, subtitles, stock quotes and jobs across X, Reddit, YouTube, GitHub, Instagram, Facebook, Xiaohongshu, Bilibili and more — no API keys for public content.
The free/local fetch layer the research cluster lacked: its fetchers are the paid ScrapeCreators API (credit-limited) and exa-search, and no entry reads platforms like Xiaohongshu and Bilibili at all. Fetch here, analyze with the cluster’s existing skills.
Evaluates a creator or brand audience from public signals: demographics, geography, language, and fit.
Prevents sponsorship and partnership picks based on follower counts alone.
Mines comments and replies for objections, questions, buying intent, and voice-of-customer language.
Comments are unfiltered market research, already written by the audience.
Researches competitors' social strategy and turns public data into a practical strategy brief.
Shows what content is winning in your niche and where the gaps are.
Turns public videos, transcripts, and posts into threads, scripts, newsletters, and content calendars.
One good source becomes a week of channel-ready material.
Tears down a creator or brand profile: positioning, pillars, hooks, formats, and outlier posts.
A repeatable way to learn from accounts that already win.
Runs autonomous multi-step research through Google's Gemini Deep Research agent and returns detailed, cited reports for market, competitive, technical, and literature questions.
The heavyweight research option: paid per run and minutes-slow, but it produces cited reports the social-data set cannot.
Calls Exa's semantic search API directly over HTTP, with ranked results, domain and freshness filters, extracted text and highlights, structured output, and streaming responses — no SDK.
A general open-web retrieval primitive for agents, complementing the social-platform research set.
Finds and scores creators in a niche for outreach, partnerships, sponsorships, and UGC.
Builds prospect lists from evidence instead of manual scrolling.
Conducts systematic, scoping, and narrative literature reviews across PubMed, arXiv, bioRxiv, and Semantic Scholar, with coverage tracking and screening, delivered as Markdown or PDF.
Brings academic-grade evidence synthesis to the research category; nothing else here searches the scholarly literature.
Finds posts that massively overperformed a creator's baseline and explains why they worked.
Modeling proven outliers is the shortest path to better content.
Validates product ideas by mining pain points, demand signals, and objections from social posts and Reddit.
Evidence-backed demand checks before building or launching anything.
Reference and access layer for the ScrapeCreators REST API: public data from 27+ social platforms.
The base skill the whole research set calls; it documents every endpoint the others use.
Produces a cited social-listening brief on what people say about a brand, topic, or niche.
A fast reputation and sentiment snapshot before campaigns or launches.
Summarizes and mines video transcripts for hooks, claims, quotes, and reusable content atoms.
Unlocks the content inside long videos without watching them all.
Finds trending topics, hashtags, sounds, and formats in a niche from public trend endpoints.
Trend timing is free reach for channels; this turns trends into content angles.
API security testing techniques and scenario playbooks for REST, GraphQL, WebSocket, and LLM-integrated APIs, mapped to the OWASP API Top 10.
The deep, scenario-level companion to security-review: the owner's bots and workers expose APIs, and this is the offensive testing reference for them.
Production-ready incident response runbook templates: severity levels, triage and escalation paths, step-by-step recovery, and communication structure.
When a live bot or worker goes down at 3 AM, a written runbook beats improvisation; this is the template set to write them from.
Secrets management for CI/CD pipelines using Vault, AWS Secrets Manager, and native platform tooling: storage, rotation, and least-privilege access.
Every project here juggles bot tokens and API keys; this is the disciplined way to store and rotate them instead of scattering them across configs.
Security checklist and patterns to run when adding authentication, handling user input, creating API endpoints, or touching secrets and payments.
The flagship security skill from the most-starred collection in this repo; turns 'did we think about security' into a concrete pre-ship checklist.
Trail of Bits' Semgrep skill: writing and running Semgrep rules for static analysis and code auditing.
The canonical static-analysis skill from a top security firm; listed link-only because its CC-BY-SA-4.0 license does not allow redistribution.
Applies the STRIDE methodology to threat-model a system — spoofing, tampering, repudiation, information disclosure, denial of service, elevation of privilege — with mitigations for each.
Threat modeling before code exists catches design flaws no later review can; this is the canonical STRIDE pattern set from a widely used plugin library.
Hardens release artifacts against AI reverse-engineering agents with decoys, analysis-budget exhaustion and dead ends, driven by the bla-agents CLI.
The defensive mirror of REA: REA takes artifacts apart; this protects your own releases from being taken apart.
Reverse engineers binaries, Electron/JS apps, .NET, firmware, and browser targets with REA, backed by runtime evidence.
The owner's reverse-engineering workhorse: explains features, compares versions, and guides reconstructions from real artifacts.
No skills match your filters. Try clearing a tag or the search.
Companion servers, not skills — recommendations only, nothing vendored. Each entry states what the server can run or access; paid services are labelled. Full list with provenance in mcps.json.
Vetted, valuable, not adopted — link-only, and not skills. Things checked and found genuinely useful that do not belong in the collection: platforms, plugins, dashboards, or infrastructure, or duplicates of an entry that already exists. Full list in void.json.