swyxio Skills

Reusable skills for Claude Code, Codex, Cursor, and similar agent environments.

Each folder is a self-contained workflow with a SKILL.md, optional supporting scripts, and a focused trigger description so an agent can pick the right tool quickly.

Use the most specific skill that matches the job. When a workflow spans multiple stages, start with the orchestrator skill, then hand off to the atomic skill for the active stage.

• Prefer API skills for stable, supported bulk operations; prefer Computer Use skills for authenticated browser states, Studio-only controls, file pickers, disabled buttons, and save verification.
• Keep source acquisition, metadata staging, and upload ledgers in batch/upload skills. Keep existing-video cleanup, thumbnails, scheduling, playlist fixes, and save recovery in youtube-studio-computer-use.
• Treat status, reviewer, and operations fields as private by default. Public descriptions should use submitted abstracts, bios, company/project links, and social links entered for publication.
• Do not physically reorganize skill folders into categories unless every target agent loader supports nested skill discovery. The top-level folder/SKILL.md layout is intentional.

Use these opt-in skills only when the named hardening problem is the primary task. They are diagnostic tools, not a maturity ladder: select the smallest relevant skill, reuse existing controls, prefer deletion, and stop when the explicit problem is resolved. Ordinary implementation work should not trigger the suite merely because it will ship or could be made more robust.

Foundation antislop-codebase — Structural cleanup/migration. Reduces demonstrated repository maintenance cost without imposing folder, file-size, compatibility, testing, or audit-artifact targets. Productization productionize-app-with-services — Bounded productization. Adds only explicitly needed product services for named users and operators; it does not install a default SaaS maturity stack. Safety security-hardening — Threat-scoped appsec. Fixes concrete reachable risks in an explicitly reviewed threat surface and reuses existing framework/provider defenses. Operability observability-hardening — Question-driven visibility. Uses the cheapest existing or new privacy-safe signal to answer named production questions without requiring every telemetry type. release-readiness-hardening — Minimal ship proof. Audits the smallest sufficient release controls, prefers authoritative provider facts, and removes redundant ceremony. vercel-production-cost-review — Material cost diagnosis. Answers a defined Vercel billing question by investigating the few dominant drivers before authorizing any remediation. Quality test-strategy-hardening — Confidence-per-minute. Solves explicit test-system problems through faithful boundaries, deterministic fixes, and deletion or consolidation—not test-layer growth.

• align-me — pauses before a long autonomous run to surface material ambiguities as numbered, lettered choices with concrete tradeoffs, recommendations, and an approve all path.
• new-mac-setup — opinionated Apple Silicon Mac bootstrap for fullstack and AI work. Installs Homebrew, shell tooling, editors, AI tools, terminal setup, and macOS defaults in a repeatable run order.
• cloudflare-production-builder — chooses among Workers, Pages, Workflows, Queues, Durable Objects, D1, R2, KV, Cache API, alarms, and Cron; then applies durable handoffs, safe caching, migration discipline, multi-tenant boundaries, observability, and live production verification.
• claude-session-introspect — inspects Claude Code session JSONL files at ~/.claude/projects/ for token totals, prompt counts, assistant turns, tool calls, compaction boundaries, and compaction summaries.
• deep-trajectory-analysis — reconstructs paired agent, game, or policy trajectories from exact shared pre-states, connects aggregate effects to first-divergence evidence, and validates machine-readable causal reports before promotion decisions.
• design-apps-with-imagegen — audits an existing interface, generates four divergent visual and behavioral directions including a wildcard, obtains one numbered approval contract, implements with code and generated assets, and compares matched mobile, tablet, and desktop captures until material deltas are resolved.
• visual-playtest — runs a concise browser playtest of visual products, verifies that controls and media paths actually work, inspects responsive scrolling and overlays, and returns severity-ranked observed defects with the smallest useful fixes.
• ceo-creator — creates a durable project CEO with an evidence model, authority boundaries, operating cadence, initiative and delegation rules, privacy protections, and a decision-oriented reporting contract.
• cassandra-creator — creates a durable independent dissent agent with separate evidence access, a dated assumptions and predictions ledger, symmetric skepticism, a high interruption threshold, explicit self-correction, read-only authority, and an optional approved recurring cadence.
• skill-cutter — critically classifies and trims overgrown skills to their behavioral core, narrows over-broad trigger metadata, and separates provider constraints from optional or project-specific policy.
• smart-entity-resolution — resolves named people or organizations in messy databases with aliases, duplicates, sparse records, common names, LLM retrieval repair, reranking, and visible runner-up candidates.
• autoreview — performs an explicitly requested final review using whatever review capability is available, without depending on a particular helper, model, or service.
• public-qa-chatbot — builds unauthenticated public Q&A chatbot widgets with rate limits, origin/input hardening, semantic caching, observability, streaming UX, and robust chat scroll behavior.
• slackbot-builder — builds production Slack bots with signed Events API handlers, causal shared-thread sessions, per-thread serialization, stateful routing and owned-resource resolution, state-aware Block Kit approvals, durable execution for slow agent work, guaranteed result-or-error delivery, and structured observability.
• sync-url-navigation — syncs URL query params with app navigation, tabs, and filter state so views are bookmarkable and shareable (view, table, q, deep links, popstate).
• app-ux-paradigms — standard web UX defaults: Esc/backdrop/× for modals, ⌘/Ctrl shortcuts, form save states, tables, menus, and help text for discoverable interactions.
• data-chatbots — copilots over structured data that propose mutations (draft → Apply), not direct writes: prompting, validator allowlists, session memory with DRAFT/APPLIED/IGNORED, version-stale UX, and test matrices.
• reserved-handle-policy — designs and implements two-tier public username protection with hard platform reservations, administrator-reviewed claims, separator-confusable matching, a source-attributed registry of common names and notable identities, and signup/rename/admin test guidance.

AI DevRel is the end-to-end publication bundle: capture technical work, acquire and transform source media, extract the useful story, publish dense written and video explanations, operate distribution channels, and measure what changed. Start with the orchestrator for a multi-stage job, then route each stage to the narrowest atomic skill.

• ai-readme — turns a repository into a progressive, executable explanation for an explicitly chosen reader, with a verified first result, one stable example, honest tradeoffs, and a context-isolated cold read.
• ai-devblog — routes technical material into the right story mode and weight, aligns on the reader and belief change, preserves primary evidence, and edits the result for clarity and human interest before verified publication.
• blog-system-design — designs dense technical blog systems: index and section pages, compact typography, full-text / search, resizable article index rails, responsive floating TOCs, reusable explanatory components, and information-dense media policy.

• media-transform — orchestrates video pipelines across download, upload, transcription, chapters, thumbnails, and title testing by routing to the right atomic skill for each stage.
• download-video — downloads embedded videos from web pages by resolving the real player URL and calling yt-dlp with the right referer/origin headers.
• download-x-video — downloads X/Twitter post videos with yt-dlp, including HLS streams and reliable final-path detection.
• zoom-download — downloads Zoom cloud recordings, verifies filenames/file types, and supports ffmpeg-based content analysis.

• transcribe-anything — transcribes audio and video files using pluggable ASR backends including local Whisper, whisperX, faster-whisper, OpenAI, Groq, Deepgram, AssemblyAI, Gemini, and Hugging Face models.
• conference-transcribe — splits long conference livestreams or YouTube videos into per-talk transcripts using chapter timestamps, segment transcription, and LLM cleanup.
• multimodal-extraction — turns local videos or video URLs into Markdown timelines with slide screenshots, key frames, and transcript spans aligned by timestamp.
• summarize-anything — recursively summarizes long text with pluggable LLM backends and can emit executive summaries, YouTube descriptions, chapters, posts, titles, thumbnail prompts, blog outlines, and pull quotes.
• podcast-publishing-assistant — turns podcasts, interviews, panels, and long-form audio/video into transcripts, summaries, chapter markers, show notes, titles, descriptions, and promo copy.

• youtube-channel-operator — designs full-power multi-channel YouTube operators with typed Data/Analytics/Reporting/Live API access, transcript-derived viewer packages, paid-versus-organic analysis, guided Slack approvals, exact-channel OAuth isolation, immutable external-action audit, Studio-only handoffs, and post-change measurement.
• youtube-api — manages YouTube videos programmatically through the YouTube Data API v3, including uploads, thumbnails, metadata updates, and channel video listing.
• youtube-publish — publishes videos on YouTube, edits titles/descriptions/timestamps, assigns playlists, and manages YouTube Studio metadata workflows.
• youtube-studio-batch-upload — batches YouTube Studio uploads from Airtable or local video submissions, with source download recovery, metadata staging, unlisted visibility, playlist tagging, save verification, and blocked-row reporting.
• youtube-studio-computer-use — automates live YouTube Studio cleanup through Chrome/Computer Use: thumbnails, schedules, playlist fixes, visibility, save-state recovery, and DOM-assisted edit pages.
• youtube-thumbnails — creates AI-generated YouTube thumbnails with prompt engineering, image generation, compression, and upload guidance.
• thumbnail-extraction — extracts interesting video frames, face crops, presentation slides, and transparent cutouts for thumbnail compositing.

• twitter-x-scraping — scrapes public Twitter/X profile and list timelines through Nitter-compatible mirror HTML, with cursor pagination, raw JSON persistence, and anti-bot fallback guidance. Does not treat public mirrors as a reliable source for a user's following graph.

• accelevents-api — reads and updates AI Engineer Europe speaker records through the Accelevents REST API while preserving full speaker payloads.
• accelevents-speaker-sync — syncs website speaker, session, schedule, room, track, and headshot changes back to Accelevents for AI Engineer Europe.
• conference-developer-endpoints — adds and reviews developer-facing conference endpoints such as llms.txt, sessions.json, speakers.json, and MCP routes.
• europe-developer-api — works with AI Engineer Europe developer endpoints, public schedule JSON, speakers JSON, MCP access, and the local aieng CLI.
• schedule-design — builds polished conference schedule views with React grids, filters, modals, favorites, sticky layouts, and normalized data.
• sync-accelevents — pulls Accelevents speaker headshots, social data, bios, and schedule metadata into local conference source data.
• testing-schedule-preview — tests the AI Engineer Europe internal Bun schedule preview and public schedule page workflows.
• web-animation-perf — debugs jank, layout thrash, and drift in JS-driven CSS animation across AI Engineer conference sites.

• One skill per top-level folder.
• Every skill must include SKILL.md.
• Add scripts only when they make the workflow more reliable or repeatable.
• Keep auxiliary docs minimal; the skill body should carry the agent-facing workflow.

Click into each folder for the detailed workflow, prerequisites, and command examples.

Clone this repo, then point Cursor at it with one symlink for the whole directory (recommended):

git clone git@github.com:swyxio/skills.git ~/Work/skills   # or any path you prefer
rm -rf ~/.cursor/skills   # only if replacing an old per-skill layout
ln -sf ~/Work/skills ~/.cursor/skills

Cursor loads each skill subfolder that contains SKILL.md. git pull in ~/Work/skills updates every skill agents see — no per-skill symlinks to add after a new commit.

Optional: link only the skills you use (omit rm above; symlink into an existing ~/.cursor/skills/ folder):

ln -sf ~/Work/skills/data-chatbots ~/.cursor/skills/data-chatbots

Do not copy SKILL.md files into ~/.cursor/skills/; always symlink to this repo.

Note: Cursor’s built-in skills live separately at ~/.cursor/skills-cursor/ — leave that alone.