Deep dives

Skill & MCP deep dives

We read the actual SKILL.md and source files of the most-installed agent skills and MCP servers, quote them line by line, and report what they really make your agent do.

Every quote is verified against the origin on the day we publish and links to the exact commit and lines it came from, or, for skills served off-registry from a vendor domain, to the canonical file and the date we verified it. Read how we write a deep dive. Ranked adoption data lives on the Top 25 leaderboard; how-to articles live in the guides. A deep dive is the third thing: what the file actually says.

36 deep dives published so far, newest first. Every quote is verified against the origin repo before publishing.

  1. Deep dive: how does lark-doc actually work?

    lark-doc is ByteDance's agent skill for reading, creating, and editing Lark/Feishu cloud documents from a coding agent, and it is unlike anything else near the top of the catalog: the file is written in Chinese, requires a vendor CLI binary, ships in ByteDance's larksuite/cli repository with a byte-identical copy on the vendor's own domain, and leads a 25-skill suite. It has 497,602 installs as of November 18, 2026.

  2. Deep dive: how does Cloudflare's security-audit skill actually work?

    security-audit is Cloudflare's agent skill for finding vulnerabilities in a codebase, with 113,479 installs as of November 18, 2026. It is built on refusals: no full audit unless asked, no finding without a crossed trust boundary, no confirmation by the agent that found the bug, and no running your code outside a sandbox.

  3. Deep dive: what does omni-auth actually make an agent do?

    omni-auth is the designated start-here agent skill of OmniRoute, the MIT AI gateway giving coding agents one OpenAI-compatible endpoint across hundreds of providers. The file is machine-generated API documentation: two credential types, optional two-factor login, an OIDC flow that now refuses sign-ins until you set an admission policy, a two-variable setup, and a capability index whose 17 rows lead to 10 sibling files. The repo holds 88,432 GitHub stars as of November 18, 2026.

  4. Deep dive: how does azure-ai actually work?

    azure-ai is Microsoft's official agent skill for Azure AI Search, Speech, OpenAI and Document Intelligence. It has 724,820 installs as of November 18, 2026 (Skillselion catalog). The SKILL.md is a routing table; the payload is fourteen one-page SDK cheat sheets covering seven Azure services across four languages, plus an offline analyzer for Speech SDK logs.

  5. Deep dive: how does azure-diagnostics actually work?

    azure-diagnostics is Microsoft's official agent skill for debugging live Azure services: App Service, Container Apps, Functions, AKS, VMs and messaging. It has 726,415 installs as of November 18, 2026 (Skillselion catalog). The SKILL.md diagnoses almost nothing itself. It is a router into a 54-file bundle, and the symptom-to-cause tables live in there.

  6. Deep dive: how does azure-deploy actually work?

    azure-deploy is Microsoft's agent skill that runs azd up, terraform apply and az deployment against a live Azure subscription, but only after a separate skill has certified the plan. It refuses to start until .azure/deployment-plan.md reads Validated, and it forbids the agent from writing that word itself. It has 724,470 installs as of November 18, 2026 (Skillselion catalog).

  7. Deep dive: how does azure-prepare actually work?

    azure-prepare is Microsoft's agent skill that turns an application into Azure infrastructure code: it writes a deployment plan, picks a recipe, generates Bicep or Terraform, then hands off without deploying anything. It has 725,509 installs as of November 18, 2026 (Skillselion catalog). The SKILL.md is 12,862 bytes of a 568,768-byte, 166-file bundle.

  8. Deep dive: how does image-edit actually work?

    image-edit is RunComfy's router agent skill for editing an existing image from a coding agent: it classifies your request, picks between Nano Banana Edit, GPT Image 2 Edit, Flux Kontext Pro and Z-Image Turbo Inpaint, then invokes the chosen model through the local RunComfy CLI. It has 447,073 installs as of November 17, 2026 (Skillselion catalog).

  9. Deep dive: how does nano-banana-edit actually work?

    nano-banana-edit is RunComfy's agent skill for editing images through one endpoint, google/nano-banana-2/edit: swap a background, remove one object, keep a face intact. It has 446,875 installs as of November 17, 2026 (Skillselion catalog). At 9,026 bytes it ranks 27th of its repo's 30 skills, and still out-documents the router that covers the same endpoint.

  10. Deep dive: how does nano-banana-2 actually work?

    nano-banana-2 is RunComfy's agent skill for generating images with Google's flash-tier Gemini image model from a coding agent. It has 447,551 installs as of November 17, 2026 (Skillselion catalog). Its first section is a routing table that decides whether you should be using this model at all, before the file explains how to run it.

  11. Deep dive: how does image-to-video actually work?

    image-to-video is a RunComfy agent skill for people who animate still images from an agent. It classifies the request, then calls one of three paid video models through the local RunComfy CLI. The bundle is a single SKILL.md with no scripts. It has 447,829 installs as of November 17, 2026 (Skillselion catalog).

  12. Deep dive: how does shadcn actually work?

    shadcn is the shadcn/ui team's agent skill for React projects that already run their component library. It shells out to the shadcn CLI for project config, routes every component lookup through the registry instead of the model's memory, and enforces 35 composition rules. It has 271,099 installs as of November 18, 2026 (Skillselion catalog).

  13. Deep dive: how does vercel-composition-patterns actually work?

    vercel-composition-patterns is Vercel's agent skill for React component APIs. It carries eight rules that trade boolean props for compound components, lifted state and explicit variants, and it triggers on refactoring and component-library work. It has 459,854 installs as of November 18, 2026 (Skillselion catalog). It is for teams whose components have started collecting flags.

  14. Deep dive: how does hyperframes-cli actually work?

    hyperframes-cli is HeyGen's agent skill for driving hyperframes, the CLI that renders HTML compositions to video, with 1,600,487 installs as of November 18, 2026. It is a 103-line command map: commands by job, a five-step loop that stops for a person before rendering, timeline edits with receipts, and project history. Its first instruction is a telemetry call.

  15. Deep dive: how does ai-research-explore actually work?

    ai-research-explore is lllllllama's agent skill for deep learning research: it holds an agent to a frozen task, dataset and evaluation, scores candidate ideas through numeric gates, and writes candidate-only evidence to explore_outputs/. It has 361,073 installs as of November 17, 2026 (Skillselion catalog). It is for researchers who already have a baseline to beat.

  16. Deep dive: how does ui-ux-pro-max actually work?

    ui-ux-pro-max is NextLevelBuilder's agent skill for interface work, and what it teaches an agent is how to run a search. The design opinions live in roughly 3MB of local data files that a bundled Python script queries; the file itself is a query manual. It has 457,320 installs as of November 18, 2026 (Skillselion catalog).

  17. Deep dive: how does remotion-best-practices actually work?

    remotion-best-practices is Remotion's official agent skill for making video with React. Its SKILL.md is 4,426 bytes of headings and conditions whose only rules of their own are about workflow, and installing it puts 1,192,223 bytes on disk, because the twelve skills it routes to ship embedded inside its folder. It has 848,164 installs as of November 18, 2026 (Skillselion catalog).

  18. Deep dive: how does emil-design-eng actually work?

    emil-design-eng is Emil Kowalski's agent skill for interface polish: a decision framework that mostly argues against animating, custom easing constants it calls stronger than the CSS defaults, and a mandated Before/After review table. It has 457,887 installs as of November 18, 2026 (Skillselion catalog). It is for frontend work where the complaint is that the result feels cheap.

  19. Deep dive: how does brainstorming actually work?

    brainstorming is obra's agent skill for finding out what you want, and why, before your agent builds it. Since v7.0.0 it does a clear task at once, asks open questions about anything else, lays out two or three approaches, and builds nothing until you approve a written design. It has 433,120 installs as of November 18, 2026 (Skillselion catalog).

  20. Deep dive: how does skill-creator actually work?

    skill-creator is Anthropic's own agent skill for writing agent skills. It specifies what belongs in the description field, prescribes an eval harness that runs every test case twice, with the skill and without it, and ships a script that tunes a description against a held-out trigger set. It has 435,334 installs as of November 18, 2026 (Skillselion catalog).

  21. Deep dive: how does web-design-guidelines actually work?

    web-design-guidelines is Vercel's agent skill for developers who want an accessibility and UX pass over UI code they already wrote. Everything it checks lives in a second repository it does not ship, behind a URL pinned to main, so a review grades your code against whatever Vercel published that morning. It has 874,105 installs as of November 18, 2026 (Skillselion catalog).

  22. Deep dive: how does setup-matt-pocock-skills actually work?

    setup-matt-pocock-skills is Matt Pocock's agent skill that configures one repo for the rest of his engineering skills. On the default path it asks one question, where your issues live (two if triage is installed), shows you a draft to approve, then records that and the domain layout in docs/agents/ and in your CLAUDE.md. It installs nothing. It has 1,158,592 installs as of November 18, 2026 (Skillselion catalog).

  23. Deep dive: how does tdd actually work?

    tdd is Matt Pocock's agent skill for building test-first: it defines what a good test is, makes the agent confirm where tests will live before writing one, and names three ways agent-written suites rot. It has 1,251,412 installs as of November 18, 2026 (Skillselion catalog). It is for anyone who has watched an agent produce passing tests that assert nothing.

  24. Deep dive: how does improve-codebase-architecture actually work?

    improve-codebase-architecture is Matt Pocock's agent skill that scans your codebase for refactors worth doing, presents them as an HTML report it opens in your browser, then interviews you through whichever one you pick. It has 1,278,323 installs as of November 18, 2026 (Skillselion catalog). Its primary deliverable is a rendered web page, not code.

  25. Deep dive: how does video-edit actually work?

    video-edit is RunComfy's agent skill for editing existing video from a coding agent: restyle a clip, swap a background, transfer motion from a reference. It has 449,661 installs as of November 17, 2026 (Skillselion catalog). Its primary job is choosing which paid model to spend your money on, and most of its links are tagged so RunComfy can attribute the click.

  26. Deep dive: how does impeccable actually work?

    impeccable is Paul Bakaus's agent skill that hands your coding agent a design director's judgment and twenty-four named commands to apply it with. It has 415,134 installs as of November 18, 2026 (Skillselion catalog) and is one of the fastest-trending skills in search. It also has a second life: a licensed derivative of its playbooks, published under a different name, which out-installed the original in our catalog in August 2026 and has since been rebuilt on a newer impeccable release.

  27. Deep dive: how does supabase-postgres-best-practices actually work?

    supabase-postgres-best-practices is Supabase's agent skill that loads Postgres rules into your agent before it writes SQL, designs a schema, or ships a migration. It has 489,355 installs as of November 18, 2026 (Skillselion catalog). The most interesting thing about it is not any single rule: it is that this file and Vercel's React skill are the same document with different subject matter, and Vercel published the shape first.

  28. Deep dive: how does design-taste-frontend actually work?

    design-taste-frontend (tasteskill 2.5.1) is an agent skill that builds whole company websites from a kit of scripts, checks them in a headless browser, and will not let the agent call a page done until the checks pass or three fix rounds run out. It has 791,993 installs as of November 18, 2026 (Skillselion catalog).

  29. Deep dive: how does caveman actually work?

    caveman is an agent skill that makes your agent answer short and plain, answer first, with every code span, number and negation kept exact. It has 3,844 installs as of November 18, 2026 (Skillselion catalog). Since October 2026 it is nine numbered rules, seven exits to full prose and a pre-send check, with its six intensity levels replaced by two separate skills for ultra and wenyan.

  30. Deep dive: how does microsoft-foundry actually work?

    microsoft-foundry is Microsoft's official agent skill for running their Foundry AI platform from a coding agent: provisioning, deploying agents, evaluating them, fine-tuning models, managing quota and RBAC. It has 735,738 installs as of November 18, 2026. It is less a skill than a platform operating manual: 231 files, twenty-four sub-skills, its own version number, and an installer script it runs the moment it loads.

  31. Deep dive: how does vercel-react-best-practices actually work?

    vercel-react-best-practices is Vercel Engineering's agent skill that turns their internal React and Next.js performance doctrine into rules your coding agent applies while it writes, reviews, or refactors code. It has 908,431 installs as of November 18, 2026. Unlike the shims we have covered in this series, this one ships the entire method: 70 rule files, a priority order, and a build pipeline that compiles and tests them.

  32. Deep dive: how does grill-with-docs actually work?

    grill-with-docs is Matt Pocock's agent skill that runs the same relentless planning interview as grill-me, and writes the documentation while you talk: a glossary of your project's terms and architecture decision records. It has 1,339,391 installs as of November 18, 2026. The installed file is seven lines long and composes two other skills to do it.

  33. Deep dive: how does agent-browser actually work?

    agent-browser is Vercel Labs' agent skill that gives your coding agent a real Chrome browser, driven from a CLI. It has 1,570,882 installs as of November 18, 2026. The file you install is a 51-line stub: the working guide ships inside the CLI's npm package, and since November it tells the agent to pass --session <name> on every command.

  34. Deep dive: how does grill-me actually work?

    grill-me is Matt Pocock's agent skill that makes your coding agent interview you relentlessly about a plan or design before it writes any code. It has 1,567,487 installs as of November 18, 2026. Here is the part almost nobody who installed it knows: the file is seven lines long, and its only instruction hands off to a sibling skill called grilling, where the whole method lives.

  35. Deep dive: how does Anthropic's frontend-design skill actually work?

    frontend-design is Anthropic's official agent skill for making AI-built web UI look designed instead of templated. It casts your agent as a design lead at a small studio, forces a token-system design plan before any code, and names five cliché traits of AI-generated design to steer away from, hex codes included. It has 1,131,985 installs as of November 2026 (Skillselion catalog); see where it ranks on the leaderboard.

  36. Deep dive: how does find-skills actually work?

    find-skills is a Vercel Labs agent skill that turns your coding agent into a skill-discovery assistant. When you ask "how do I do X", the agent searches the open skills ecosystem with the npx skills CLI, vets the results by install count and source reputation, and offers to install the best match. It has 4,410,475 installs as of November 2026 (Skillselion catalog), which makes it the most-installed agent skill in the Skillselion catalog.

This week in AI coding

Five minutes, every Monday - the tools, releases and tactics for developers.

unsubscribe anytime.