An Agent Resource Collection: Skills, MCP Servers, Plugins, and Handy Tools

This English version was machine-translated from the Chinese original. Technical terms may require verification.

This article is a collection of resources that I like and can reach out to Agent/ComingCLI.

The questions in this article can also be addressedMCP (Model Context Protocol)、From MCP to Argentina Skills: Why does Agent need a new context work protocol?How the concept of a relatively close read together is developed in different contexts.

They are not necessarily large, nor do they warrant a separate article. It could be a Skills, a MCP Server, a plugin, or a script, a hint structure, a little bit strange, but really saves things. Common ground: it can be hung on Agent or Coding CLI, and it can be a little more powerful.

This is a two-part piece: one is.Large plant / Platformr Basic Services SkillCLI, product files and engineering best practices maintained by official or platform teams; the other isSomething that was invented by independent developers.More like a test field, there are editing processes, creative workflows and small and easy tools. Each item is still a standard type (Skill / MCP Server/ Plugin/ Tool) and will not be able to be saved half a day later.

Stay here. We'll get new ones later.

Large plant / Platformr Basic Services Skill

vercel-labs/skills

Type: Skill Distribution Tool / CLI

vercel-labs/skills Not a specific skill, but a tool to play skill. The entrance is... npx skills, you can pull the skills from GitHub repository, local directory or other guit source, and place them in different locations that Agent can identify. It doesn't teach Agent how to write pages, make videos, run experiments, but one thing: how to fit, how to share, how to migrate.

It's also about to be with vercel-labs/agent-skills Distinct.vercel-labs/skills is a CLI/ Distribution Tool, like Infrastructure;vercel-labs/agent-skills The only content collection that is officially maintained by Vercel is the Skills collection. The first one does what it does, the second one does what it does. The two warehouses were together, and it became clear that Vercel wasn't just going to write a few good ones. SKILL.mdI also want to make this distribution easier. Vercel also wrote some blog posts about their research on Skills, and their views are worth reference.

ClawHub

Type: Skill / Plugin registration form / distribution directory

ClawHub is public space in OpenClaw ecology. It's surrounding. SKILL.md The publication, version, search, installation, comment, collection and review of the documents and the accompanying documents are included in the same catalogue. The specific tasks are still being given to individual skills; ClawHub is concerned about how these skills and plugins are being discovered, updated and governed.

Put it in the... vercel-labs/skills It'll be clearer if we put it together.vercel-labs/skills Installation and migration: Put the skill in the right place from GitHub, local directory or guit source. ClawHub favours OpenClaw's own registry and package datalog: both skills catalogues and the start of covering more heavy extension units such as code plusins, and the bookle pluss. For OpenClaw, it is the public market and indexing layer.

When using such forms, do not simply read names. Before installation, inspect: Source, version, Changelog, metadata, scanning status and moderation status are worth looking at. The Skill threshold is a good thing, but it brings scripts, environmental variables, external services and local privileges together into the Agent workflow. The more you get, the better you get to see it first.

Remotion Agent Skills

Type: Skill

Remotion Agent Skills is a video project statement prepared by Remotion for Agent. It allows Agent to write React respect, frame-based animation, audio, subtitles, material and transpositions in a Remotion style, and to end up with MP4. The time here is not one page of advance, but a frame of calculation; the product is not a web stage, but a video file that can be delivered, reused, automatically exported.

It pulls back the video to the engineering floor. Many AI video tools are suitable for output atmosphere and lens (auto-matching of photographic material), but technical content, data animation, subtitle rhythm, batch templates and CI rendering still require more definite structures. Remotion makes Agen no longer imagine a video, but rather use React and the timeline to make it. It's really about making a stable video, or mass render in the CI, more like a production engine. We still need to rely on these certainty tools before the cost of video generation models is reduced.

Something that was invented by independent developers.

General Research Review Agent And Skills

Type: Skill

This is a set of research reports that I made, mainly to assist with general writing. It's not a single line of "help me write an overview" but it's a few steps to break the scooping review down: finding literature, weighting, screening candidate pools, extracting paper evidence, organizing outlines, consolidating the text, and finally checking references and language.

My favorite thing is that it turns the synthesis from a chat to a process with intermediate products. In the past, such work was often done by Zotero, Google Scholar, manual forms, scattered PDF, temporary prompts, and finally LLM to help with the tanning. It can certainly be done, but it's painful to look back: which articles are excluded, why, which ones are supported by, and whether or not the quotations are mistook.

This skills handles these places. The final PDF/HTML/Markdown report will be left. It is not about making the article more beautiful, but about making the most easily dispersed part of the overview available: where it comes from, where it filters, where it is written, what it supports. It is more suitable for a broad-based, audit-required study than an ad hoc prompt or simple documentation management tool; if only a few papers were summed up, there would be no need for such a process.

Humanizer / humanizer-zh-next

Type: Skill

Humanizer and humanizer-zh-next are a group of writing clean-up skills to change the AI taste in text. They are not a grammatical correction tool, but they are more annoying: over-summation, propaganda cavity, tri-parte ranking, abuse of a dash, a broad “key role”, and a kind of a charade that looks like a chat robot answering. Humanizer-zh-next was a new Chinese version of the Skill, which I developed and maintained, which was updated and enhanced on similar features.

I like them because it's more specific than "do me a favor." Normal tweaking often makes text more smooth and full, but it's more like AI; Humanizer, in turn, deletes too much of the sort of place that is too neat, too much of a summary, too much of a template. The English text can be found in Humanizer, and the Chinese text can be used as humanizer-zh-next.

Humanizer-zh-next is particularly useful for Chinese bloggers. The Chinese model is easy to write “first, second, and above” and to take a general judgement to “important” “deeply affect” “complex patterns”. I don't want to make it more gorgeous, but I want to make it sound like someone who really wrote, deleted, hesitated. It can be used once the technical articles, project presentations, comments and essays have been completed.

Beautiful Article

Type: Skill

Beauutiful Article is a skill-format of ready-made materials into a web site. Web pages, PDF, Word, Markdown, screenshots and even a large amount of paste material can be handled by it first. It will make these things clean, and then it will make a single document that can be opened offline and shared directly. It's not about creating a web page, but about making Agent really like an editor, reading material first, reordering it, making it better read, and giving us a more elegant interactive way of doing things.

Beauty is not a stack of styles, and for an article, better reading requires a combination of content and visual structure. It draws material before writing editing plans, and then stops to allow users to confirm the type of articles, themes, layouts, layouts and cover. It's not a matter of breath, but of first-screening and first-sectioning, right? This rhythm is very good for long materials: style can be calibrated early and information can not easily be lost quietly during generation. The intermediate files will remain and will not be left with only one finished HTML when checked back.

It is appropriate to deal with material that is actually good, but that is read too painful, such as reports, tutorials, interviews, flashbacks, programme analysis, explanatory texts and long articles, and can also be used for personal sharing. The final product can be a table, code block, diagram, cover, directory, or a little interactive section, but it is always an article, not a backstage, form, dashboard or product prototype. If there was one thing in hand that was solid enough, it was too long, too loose, too hard to read, I would be happy to give it to Beauitiful Article to clean it up.

Web Video Presentation

Type: Skill

Web Video Production is following the subject of content directorship. It starts by tearing articles or oral articles into scripts, outlines and chapters step, and then making a 16:09 web stage that can be pushed by click and fit for screen. It's not responsible for rendering the React code MP4, it's concerned about how to organize an article into a speech worth recording.

It does not come up with animated drawings, but requires that the original text be retained, that the text be produced and outlined, and that the text be aligned with the text, chapters, themes, material, development models. It's... narrations.ts The only source of truth is the number of step and the text of the interview, which is taken from the screen at every step, and the details of the video are taken back into the article. This idea is particularly useful for tech bloggers: technology content is the most dangerous to be put on a few pretty pieces of crap, and it forces Agent to leave the chain of argument, numbers, cases and rhythm behind.

So when you do a single technology sharing, it goes more than straight up Remotion: first, you break your blog down into a blog, chapter and visual beat, then you get a video that is sufficiently clear on the browser. If more stable subtitles, audio and batch export are needed later, it is not too late to move these intermediates to the Remotion for final rendering. Web Video Production offers a good layer of content and collaborative rhythm, especially for making articles a narrated, logged and detailed web presentation. Nor is it necessary to bear the high cost of video generation and the risk of a card.

yinyo-image2-prompt / GPT Image 2 Skill

Type: Skill

Both skills revolve around the same question: how to get Agent to use GPT Image 2 drawings and redraws more steadily than a temporary set of hints. The new generation of image models is not taking much of the old spell like “8K, ultra detailed, masterpiece”, and the real effect is on the main order, the scene structure, the text constraints, the lens language, what changes and what cannot be changed when editing. They deal with this level of translation: turning people's vague ideas into visual instructions that models can more easily implement.

Yinyo-image2-prompt is a small specialized. It has broken down the common GPT Image 2 tasks into 35 sub-templors, using a 5-Phase process to guide users to the first-level assessment of the scene and then to fill in key parameters, and behind it has been a compilation of 33 sets of blind reviews and several sets of experience with the PK. It reminds you of when not to put templates: App UI, Logo, editorial workflow, structural tasks, which are helpful; Pixar 3D, conceptual illustrations, which are more narrative and graphic tasks, over-engineering can overwhelm the model ' s freedom. This is more useful than just a few prompts.

The GPT Image 2 Skill of ConardLi is more like a biographic production line. It is not just a prompt, but it also divides the operating environment into three modes: When you have an API key locally, you can call the script and drop the disc; when you bring the host Agent with the image tool, you entrust the rendered prompt to the host; if you have no image tool, it can be degraded into a pure prompt consultant. It's also got... check-mode.js、generate.js、edit.js These scripts, which are run by directories like UI, product, information diagram, academic chart, technical chart, editing workflow, etc., are also archived separately and are suitable for repeated use in the actual Agent workflow.

The yinyo-image2-prompt is better suited to the hints used to learn and optimize GPT Image 2: What to ask, what to ask the province, and what to overwrite. GPT Image 2 Skill is more suitable as an entry point for the project: Agent is given first to judge the mode, then to select templates, retrofitt, save the product and, if necessary, to make a direct drawing. The former are like brains, the latter are like hands and feet; when one really wants to make a living map a stable stream of work, it's more complete to put them together.

It's a game. It's a good idea to use it for money.