Best Skills for Autonomous Agents

Escape generic AI-generated UIs. Forces a bold, distinctive design direction before any code is written.

One skill covering every document format the category exists for — including scanned PDFs and RAG chunking — from the maintainers of the parser itself.

Write HTML, render video. Agent-native, deterministic, no build step.

Structured multi-step development. Brainstorm → spec → plan → build → review → merge — TDD baked in.

Microsoft's first-party Playwright skill, invoked through the CLI — the reference implementation for driving a real browser, with no credentials required.

Diff-aware adversarial QA: it reads the git diff and tests only what changed, across correctness, accessibility, layout and UX.

Anthropic's own sales stack: research → outreach → call prep → pipeline, nine skills in one plugin.

Stripe's own integration playbook: pick the right API, set up Connect and billing, and handle keys and webhooks securely.

The marketplace architecture decision, made with a compatibility matrix that blocks invalid Connect configurations instead of describing them.

Chrome's own team, giving agents the DevTools protocol: drive it, inspect it, debug it.

The only vendor-official skill from a major diagramming product — and it emits editable .drawio files rather than a throwaway image.

Vendor-official voiceover generation across 70+ languages, from the company that defined the category. Ships with evals, which almost no skill does.

Transcription and subtitle generation as a first-class skill — the media job agents are asked for most and could not previously do well.

A self-updating vector-database skill: it fetches current Qdrant guidance at run time instead of shipping a snapshot that rots.

Local document extraction with no API key — read the file, not the whole file.

Hookdeck's webhook playbook: 40 provider-specific skills covering signature verification, idempotency, and event handling — each with runnable examples.

Composio's official skill for connecting agents to 1000+ external apps: auth configs, per-user connections, tool execution, and triggers.

1,000+ SaaS integrations as one skill. OAuth lifecycle, scoped credentials, structured actions.

Supabase done right. RLS policies, edge functions, realtime, auth — in idiomatic patterns.

The generalist Postgres skill the ecosystem was missing — schema, indexing, extensions and migrations, from a Postgres company.

Zero-downtime Postgres migrations, with the locking behaviour spelled out — which ALTER TABLE takes an AccessExclusiveLock and what to do instead.

Schema design decisions made explicitly: data types, key strategy, JSONB versus columns, and when partitioning earns its complexity.

First-party HubSpot CRM operations for agents: lookups, deals, quote-to-cash, reporting.

Full-stack WorkOS implementation with auth provider migration paths.

Drop-in WorkOS widgets with backend token wiring across every major framework.

Web scraping, search, and browser automation built to be agent-friendly.

Paddle's own integration playbook: ten skills for Checkout, webhooks, pricing pages, and subscription lifecycle on the merchant-of-record platform.

The SQLite slot, filled by Turso — and unusually, a skill that actively steers the agent off the stale API surface web search returns.

Slack's own Block Kit skill — and the rare vendor skill that needs no token, because it validates against Slack's public API.

OpenAI's chart-design skill: the judgement layer that decides which chart the question actually needs before any plotting code is written.

Dashboards with source definitions and a QA step — the parts that separate a monitoring view from a wall of charts nobody trusts.

Extract, fill, merge. PDF manipulation with bundled scripts so Claude does not waste tokens generating them.

Real exploits, no false positives. 96.15% exploit success across 50+ vuln types.

A vendor-maintained rule corpus for secure code review — one reference file per vulnerability class, from SQL injection to Terraform misconfiguration.

Pre-handoff a11y audit that catches what design QA misses.

Deterministic web control. ref-based clicks/fills, no flaky selectors, parallel sessions.

All of Railway in one skill: provision, deploy, debug — listed in the official plugin marketplace.

ADR authoring + design review. Trade-offs, consequences, alternatives.

Disciplined debugging loop. Reproduce, isolate, diagnose, fix.

The pre-ship checklist that catches the embarrassing stuff.

Technical writing that engineers will actually read and update.

Incident triage + comms + blameless postmortem in one harness.

System design from requirements. APIs, data, boundaries, scale.

Honest tech-debt audit with a prioritized cleanup backlog.

Layered test plan sized to the change's risk profile.

Review for reuse, quality, efficiency. Catches over-abstraction and dead code before merge.

FHIR to prior-auth to ICD-10 coding: the official payer/provider/pharma toolkit.

One-command productivity bootstrap. Memory, tasks, dashboard.

Cited research reports from fanned-out web search with adversarial verification.

Memory hygiene. Merge, fix, prune.

Real feature verification — run the app, drive it, prove it works.

Bootstrap the right plugin stack for your role in one guided pass.

100+ accessibility and UX rules from Vercel, applied as a structural audit on your UI code.

57 React/Next.js performance rules, applied as a structural review on your code.

Security audits run by a real security firm. CodeQL + Semgrep + audit workflows.

OWASP-published review, SCA, secrets and threat-modelling skills — the most citable provenance available for security review vocabulary.

Senior-designer feedback on demand. Hierarchy, usability, consistency, motion.

Turn a final design into a spec engineers can implement directly.

Design-system hygiene at speed. Audit, document, extend.

Microcopy that respects the user. Errors, empty states, CTAs, onboarding.

Always-current Claude API reference and SDK docs. Bundled with Claude Code.

Web search + 36 authoritative data sources. SEC, PubMed, ChEMBL, FRED — agent-grade research.

The official entry point for biomedical research workflows in Claude Code.

One-command app launch tailored to your project type.

Pre-merge security sweep on the current branch's diff.

The canonical Anthropic code review. Effort levels from quick to ultra.

Bootstrap a CLAUDE.md so future sessions start informed.

PR review with inline GitHub comments via gh.

Edit settings.json safely. Hooks, perms, env, behaviors.

Rebind keys and add chords without reading the keymap docs.

Cuts the permission-prompt friction by learning your safe-tool patterns.

Recurring task runner. Self-paced or interval-based.

Cron for remote Claude Code agents. Set it and forget it.

20 SEO + GEO skills in one bundle. Keyword research, audits, rank tracking, AI search optimization.

Turn raw research into themes, segments, and a prioritized next-steps list.

Research-ops in a skill. Plan the study, run it, synthesize the findings.

Brand-voice QA before content ships. Severity-ranked findings with rewrites.

Launch-ready campaign brief in one pass. Calendar, channels, metrics included.

Battlecards on demand. Positioning gaps, content holes, threat assessment.

Channel-correct content with SEO and CTA awareness baked in.

Channel-specific draft + headline options. SEO-aware out of the box.

End-to-end drip campaigns. Copy + timing + branching + benchmarks.

Executive-ready performance reports with next-period priorities.

Full SEO health check with quick wins vs strategic investments split.

Workplace memory that actually understands your team's shorthand.

Lightweight task tracking in a single markdown file.

Periodic memory + task refresh from real activity.

TDD with red/green/refactor as a concrete TypeScript walkthrough.

Structured PR reviews with severity-tagged findings — bugs, security, perf, style.

Tailor a plugin to your org's stack in one pass.

Scaffold a new plugin from scratch with guided prompts.

Terraform expertise on tap. State inspection, module validators, safety-first IaC.

Astro from scaffold to live site — content collections, search, i18n, multi-host deploy.

Multi-agent orchestration. Spawn parallel subagents and merge their outputs reliably.

Operate a Lemon Squeezy store from the agent: customer lookup, refunds, discount codes, receipts, and a sales dashboard — six support-and-revops skills.