Global Top 5 Agent Skills: Interpreting Rankings and Validation Criteria ======================================================================== This article analyzes the purposes and differences of five agent skill repositories presented as top-ranked in materials dated August 10, 2026. Because there is no reproducible GitHub snapshot for the displayed rankings and figures, they are not definitive statistics. Repository popularity must also be distinguished from the quality of individual skills. - An Agent Skill is an executable knowledge package that bundles instructions, scripts, and reference materials in a single folder for an agent to load when needed for a specific task. - The Top 5 in the provided materials are obra/superpowers, affaan-m/everything-claude-code, mattpocock/skills, multica-ai/andrej-karpathy-skills, and anthropics/skills, in that order, but they must be independently verified against historical GitHub data. - Because GitHub stars are awarded to an entire repository, they do not directly indicate the installation count, execution success rate, or maintenance level of individual skills. - The leading repositories share a focus on controlling work processes, such as clarifying requirements, planning, testing, and review, rather than immediately generating results. - Before actual adoption, licenses, prompt injection, script permissions, external communications, and version pinning must be checked. An AI agent's skill (Agent Skill) is not simply a collection of prompts. It is an executable knowledge package that bundles the instructions, executable files, and reference materials needed to perform a specific task, allowing the agent to discover and read them when needed. This document analyzes ranking data provided as of August 10, 2026, but does not treat the displayed figures as confirmed GitHub statistics. GitHub stars and installation counts are different metrics, and verifying figures from a past point in time requires the API responses or preserved snapshots from that time. What Is an Agent Skill? An Agent Skill generally contains the following elements within a single directory. Metadata: Indicates the skill's name, purpose, and conditions for use. Core instructions: Usually records the workflow, constraints, and verification criteria in SKILL.md. Scripts: Executes repeatable processes or processes requiring accuracy through code. Reference materials: Provides API specifications, organizational rules, examples, templates, and more. Result validation criteria: Specifies tests, reviews, and completion conditions. The key is not to include every instruction in the context each time, but to let the agent find skills relevant to the task and progressively read only what it needs. This structure reduces the burden of long system prompts and makes it possible to reuse the same procedures across multiple tasks. Differences Between Skills and Prompts Category General prompt Agent Skill Basic unit A single request or conversational instruction A package composed of directories and files When used When entered directly by the user Can be selected dynamically for relevant tasks Contents Mainly natural-language instructions Instructions, scripts, materials, templates Reusability Depends on the conversation and user habits Can be version-controlled in a repository Validation method Results are checked afterward Can embed planning, testing, and review procedures How to Read the Top 5 Rankings and Figures The provided material presents the following five repositories and figures as a ranking by GitHub stars. However, it does not include date-specific API results or preserved snapshots substantiating these figures. Therefore, the numbers in the table below are figures stated in the provided material, not independently verified statistics. Provided rank Repository Figure stated in the provided material Primary nature 1 obra/superpowers 269,762 A collection of skills that controls the development process from requirements confirmation through testing and review 2 affaan-m/everything-claude-code 239,034 A broad collection of agents, commands, skills, hooks, and other components for Claude Code 3 mattpocock/skills 211,297 A collection of development skills that uses questions to refine ideas and designs 4 multica-ai/andrej-karpathy-skills 200,937 A community-created collection of instructions emphasizing the connection between the reason for a change and the request 5 anthropics/skills 167,251 Anthropic's public skill examples for use in documentation, development work, and more Why This Cannot Be Called an Exact Download Ranking GitHub stars are not installations. They can also indicate interest or serve as bookmarks, and do not show whether something was actually run. Stars are measured per repository. Even if one repository contains dozens of skills, they do not reveal which skill is popular. Current values do not prove past values. Reproducing the ranking as of August 10, 2026 requires API responses saved that day or a reliable archive. Results vary depending on the scope of aggregation. It is first necessary to define whether repositories, individual directories, package installations, or execution calls are being counted. Forks and copies may be mixed in. If the same skill is duplicated across multiple harnesses or translations, it is difficult to compare scale based on file counts alone. It is therefore safer to view this ranking not as the “confirmed global Top 5 by usage on August 10, 2026,” but as an exploratory list for comparing the provided candidate repositories. No. 1 Candidate: obra/superpowers The core value of obra/superpowers lies in controlling the workflow so that a coding agent does not immediately jump into implementation. It encourages the agent to confirm requirements, organize the design, divide the plan into small tasks, and then complete the work through testing and review. Its main features are as follows. Confirms the user's actual purpose and constraints before implementation. Isolates changes by using a separate branch or workspace. Breaks large requests down into small, verifiable steps. Emphasizes test-first development and systematic debugging procedures. Rechecks test results and the scope of changes before declaring completion. An important design principle demonstrated by this repository is that “preventing premature implementation” may be more important to agent quality than “generating more code.” No. 2 Candidate: affaan-m/everything-claude-code The affaan-m/ECC in the provided material is interpreted as referring to the public repository affaan-m/everything-claude-code. Rather than being a single skill, this repository is closer to a toolbox that broadly collects various components needed to operate Claude Code. Components generally worth examining include the following. Agents and subagents for different roles Commands and skills that invoke repetitive tasks Hooks that intervene before and after tool execution Project rules and context management methods Examples of integrations with external tools or MCP Its broad range of components is advantageous for rapid experimentation, but copying every file at once is difficult to recommend. Because hooks and scripts may be involved in modifying files or executing commands, only the necessary items should be selected, reviewed, and then adopted. No. 3 Candidate: mattpocock/skills mattpocock/skills is a repository that publicly shares the agent instructions used by Matt Pocock, who is known as a TypeScript educator. A representative approach is to have the agent avoid accepting the user's idea as-is and instead ask repeated questions to reveal ambiguities and design decisions. This approach is useful in the following situations. When feature requirements are provided in only one or two sentences When an implementation method comes to mind but the problem definition is unclear When design decisions must be documented before writing code When explicit reasoning is needed for team members to review The license must not be inferred as a single license covering the entire repository. The notices in individual directories and files must be checked in addition to the root license, and whether modification, redistribution, and commercial use are permitted must be assessed separately. “Publicly available source” and “open source” do not mean the same thing. No. 4 Candidate: multica-ai/andrej-karpathy-skills multica-ai/andrej-karpathy-skills is a community repository created based on Andrej Karpathy's name and publicly known development principles. Even if a person's name appears in the title, it should not automatically be interpreted as an official project created or endorsed directly by that person. A noteworthy perspective in this repository is the traceability that connects code changes to user requests. What request makes this change necessary? Were files modified even though they did not need to be changed? Does abstraction or refactoring obscure the original purpose? Can each change be explained through a test or an observable result? This approach helps reduce the problem of agents making excessive modifications beyond the scope of the request. However, rather than claiming that it fully reproduces a specific expert's way of thinking, it is more accurate to view it as a reconstruction of public principles into agent instructions. No. 5 Candidate: anthropics/skills anthropics/skills is an Agent Skills repository published by the Anthropic organization. It can serve as a reference point for understanding the folder structure of an Agent Skill and how to write SKILL.md, as well as for examining practical examples. The number of SKILL.md files in the repository changes depending on the branch and commit. Including translations, test fixtures, harness-specific copies, and examples can inflate the count, so stating the number of files at a particular point in time requires the following conditions. Target commit hash Branch searched File search command used Rules for excluding translations and copies Directory criteria used to identify canonical versions Therefore, the figures in the provided material—“897 total, 284 canonical”—are difficult to cite as confirmed figures until the calculation procedure and commit have been verified. Common Design Principles Found in Popular Repositories 1. They Prioritize Controllable Procedures Over Autonomy Good skills do not simply give agents more authority. They establish gates such as requirements confirmation, plan approval, testing, and review to prevent agents from moving rapidly in the wrong direction. 2. They Structure Experts' Decision-Making Processes Rather Than Their Results Useful skills do not stop at role-playing instructions such as “answer like an expert.” They turn into executable procedures the questions experts ask first, the evidence on which they base decisions, and when they question the results. 3. They Confirm Why Before How The top candidates confirm the purpose in different ways. superpowers: Confirms the actual requirements and design before implementation. mattpocock/skills: Reveals gaps in an idea through questions. andrej-karpathy-skills: Traces code changes back to the original user request. This shows that the bottleneck for coding agents lies not only in syntax or code-generation ability, but also in clearly defining what should be changed and why. 4. They Specify Completion Conditions A task is not complete simply because “the code was written.” Observable completion conditions are required, such as passing tests, meeting requirements, confirming the scope of changes, and completing a review. What Rankings Miss: Security, Reproducibility, and Maintenance Popularity rankings alone cannot determine whether a skill is safe and practical. In actual adoption, the following factors may be more important than star counts. Review area Question to check Example risk Permissions Does the skill access the shell, file system, or network? Unnecessary file deletion or external transmission Prompt injection Is it designed to trust instructions in external documents? A web document overrides existing rules Supply chain Does it immediately run remote scripts or packages? Execution of modified dependency code Secrets Do logs and results include tokens or environment variables? An API key is printed or committed Version Is it pinned to a specific commit? Behavior and permissions change after an update License Are the conditions for use, modification, and distribution clear? Violation of restrictions on commercial reuse Maintenance Are recent changes and issue responses ongoing? Outdated tool calls and vulnerable examples Evaluation Has the success rate been measured across repeated runs? Only impressive examples succeed while general tasks fail In particular, a skill's natural-language instructions can have effects similar to executable code. This is because, when an agent has permission to use tools, a single line of instructions may lead to command execution or file modification. How to Create a Trustworthy Popularity Ranking To create a reproducible ranking, the following information must also be disclosed. Measurement time: Record not only the date but also the time zone. Target list: Disclose the repositories to include or exclude and the criteria for handling forks. Metric definition: Specify whether GitHub stars, forks, installations, or executions are being measured. Raw responses: Preserve the GitHub API's JSON results and the collection script. Commit pinning: Use commit hashes for file counts and license analysis. Deduplication: Establish rules for handling translations, copies, and vendor directories. Parallel quality metrics: Also consider test pass rates, issue responses, recent releases, and security review results. A more useful evaluation presents multiple metrics separately rather than relying on star count alone. For example, “interest” can be measured by stars, “adoption” by installations, “activity” by recent contributions and issue responses, and “quality” by success rates in repeated evaluations. Checklist for Choosing Skills Yourself Do the task you want to solve and the skill's invocation conditions align? Have you distinguished risky operations that require user confirmation before execution? Have you personally reviewed the contents of scripts and hooks? Have you checked external network requests and the destinations of data transfers? Does the license permit the organization's intended use? Have you adopted only the necessary files rather than the entire repository? Have you pinned the commit or release version? Have you repeatedly tested it with evaluation tasks similar to the actual project? Have you used an isolated environment where changes can be rolled back if it fails? Conclusion The five repositories presented in the provided material illustrate different directions in the agent skill ecosystem. superpowers excels at procedural control, everything-claude-code at broad operational configuration, mattpocock/skills at question-driven design refinement, andrej-karpathy-skills at tracing the reasons for changes, and anthropics/skills at understanding skill structures and examples. However, repository star counts should not be interpreted as the global popularity or quality of individual skills. Actual selection should prioritize suitability for the task, permission scope, licensing, security, maintenance status, and repeated evaluation results over rankings. FAQ Q. How are agent skills different from general prompts? A. General prompts are often instructions given directly in a conversation, whereas agent skills are packages that bundle instructions, scripts, and reference materials into directories so they can be reused and version-controlled. An agent can select skills relevant to a task and read the necessary content progressively. Q. Is the skill with the most GitHub stars the best one? A. No. Stars only indicate interest in a repository and do not guarantee actual installations, execution success rates, security, or the quality of individual skills. If a repository contains multiple skills, it is also impossible to know which skills are actually being used. Q. Is the Top 5 ranking in this document a definitive global ranking? A. No. It is an analysis of the order listed in materials provided by the operator and dated August 10, 2026, and the figures were not independently confirmed because no GitHub API response or preserved snapshot from that time was available. It is appropriate to interpret it as a list of candidates for exploration. Q. Can past rankings be verified using the current number of GitHub stars? A. Because the current repository page shows current values, it cannot directly prove the values on a specific date in the past. Verifying past rankings requires an API response saved at the time, collection logs, or a reliable archive. Q. Can publicly available agent skills be freely used commercially? A. Public availability alone does not permit commercial use. You must check the license at the repository root and any separate notices for individual files or directories, and review the conditions for modification, redistribution, and commercial use separately. Q. Do agent skills also require a security review? A. Yes. Skills may instruct an agent to invoke shell commands, the file system, external networks, or development tools. Scripts and hooks, external communications, handling of secrets, and dependency installation methods should be reviewed before testing them in an isolated environment. Q. Is it better if a repository has more SKILL.md files? A. File count is not a quality metric. It may include translations, test fixtures, templates, and duplicates, and some skills may conflict with one another or be unmaintained. You should fix the commit and counting rules and evaluate only the canonical versions that are actually needed. Q. Is a skill named after an expert an official project by that expert? A. Not necessarily. It may be a repository in which the community has reconstructed public statements or working principles. You should check the repository owner, notices in the README, and official organization accounts to distinguish the relationship to its creation and approval. Q. What is the most important characteristic of a good agent skill? A. It is important to first confirm the task's objective and scope, and to clearly distinguish planning, execution, testing, review, and completion criteria. Risky tool calls should require user confirmation, and it should be possible to roll back if they fail. Q. Can multiple skills be installed at once? A. It is difficult to recommend. Rules from different skills may conflict, and unnecessary hooks or scripts may broaden the scope of permissions. It is safer to select only the necessary functionality, pin versions, evaluate each skill, and then combine them. Sources - obra/superpowers GitHub repository: https://github.com/obra/superpowers - affaan-m/everything-claude-code GitHub repository: https://github.com/affaan-m/everything-claude-code - mattpocock/skills GitHub repository: https://github.com/mattpocock/skills - multica-ai/andrej-karpathy-skills GitHub repository: https://github.com/multica-ai/andrej-karpathy-skills - anthropics/skills GitHub repository: https://github.com/anthropics/skills - GitHub REST API documentation: Get a repository: https://docs.github.com/en/rest/repos/repos?apiVersion=2022-11-28#get-a-repository Images - Five rated skill folders on podiums with a magnifying glass inspecting a frozen entry: https://injoys.com/rails/active_storage/blobs/proxy/eyJfcmFpbHMiOnsiZGF0YSI6Nzg4NiwicHVyIjoiYmxvYl9pZCJ9fQ==--70b2edce4772f738c271f7bd778558fdecc6f4ca/ai-b8447c98.webp - Five-step validation flow from questioning to structuring, testing, searching, and approval, with safety icons: https://injoys.com/rails/active_storage/blobs/proxy/eyJfcmFpbHMiOnsiZGF0YSI6Nzg5MiwicHVyIjoiYmxvYl9pZCJ9fQ==--32bf57efa57a23995bc048d40d235f6bc81aa9e7/ai-4d2d60cd.webp --- Category: Report Source: https://injoys.com/en/articles/global-agent-skills-top-5-ranking-methodology License: cc_by Translation-Status: reviewed