{ "query": "", "kind": "skill", "page": 1, "pages": 2, "pageSize": 24, "total": 35, "results": [ { "kind": "skill", "slug": "ai-research-reproduction", "name": "AI Research Reproduction", "vendor": "lllllllama", "description": "Orchestrates end-to-end reproduction of a deep-learning repository — reading the README, setting up the environment, running training or evaluation, and tracking evidence of what actually ran.\n\nCapabilities are filesystem read and write, shell execution, and outbound network: the full-control combination, and correctly so. Reproducing research means executing code from a repository you did not write, with dependencies you did not audit, which is one of the most direct paths from \"an agent read something\" to \"arbitrary code ran on this machine\". Container it.", "website": "https://github.com/lllllllama", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "filesystem:read", "filesystem:write", "shell:execute", "network:outbound" ], "packaging": null, "configFiles": null, "category": "research", "updatedAt": "2026-09-02", "risk": { "score": 100, "band": "critical", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 }, { "id": "high-risk-capabilities", "kind": "capability", "label": "2 high-risk capabilities: filesystem:write, shell:execute", "contribution": 20 }, { "id": "exfiltration-risk", "kind": "combination", "label": "Data exfiltration risk (high)", "description": "Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.", "contribution": 15 }, { "id": "filesystem-write-shell-risk", "kind": "combination", "label": "Persistence + execution risk (medium)", "description": "Shell execution plus filesystem write means the agent can plant persistent backdoors (e.g. modifying startup scripts).", "contribution": 5 }, { "id": "full-control-risk", "kind": "combination", "label": "Full-control risk (critical)", "description": "Shell + filesystem write + network is effectively a remote shell on the employee machine.", "contribution": 25 } ], "dangerousCombos": [ { "id": "exfiltration-risk", "name": "Data exfiltration risk", "description": "Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.", "capabilities": [ "shell:execute", "network:outbound" ], "severity": "high" }, { "id": "filesystem-write-shell-risk", "name": "Persistence + execution risk", "description": "Shell execution plus filesystem write means the agent can plant persistent backdoors (e.g. modifying startup scripts).", "capabilities": [ "shell:execute", "filesystem:write" ], "severity": "medium" }, { "id": "full-control-risk", "name": "Full-control risk", "description": "Shell + filesystem write + network is effectively a remote shell on the employee machine.", "capabilities": [ "shell:execute", "filesystem:write", "network:outbound" ], "severity": "critical" } ] }, "dangerousCombos": [ { "id": "exfiltration-risk", "name": "Data exfiltration risk", "description": "Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.", "capabilities": [ "shell:execute", "network:outbound" ], "severity": "high" }, { "id": "filesystem-write-shell-risk", "name": "Persistence + execution risk", "description": "Shell execution plus filesystem write means the agent can plant persistent backdoors (e.g. modifying startup scripts).", "capabilities": [ "shell:execute", "filesystem:write" ], "severity": "medium" }, { "id": "full-control-risk", "name": "Full-control risk", "description": "Shell + filesystem write + network is effectively a remote shell on the employee machine.", "capabilities": [ "shell:execute", "filesystem:write", "network:outbound" ], "severity": "critical" } ] }, { "kind": "skill", "slug": "browser-act", "name": "Browser Act", "vendor": "browser-act", "description": "A CLI for browser automation that includes stealth features and CAPTCHA solving, aimed at driving sites that actively resist automation.\n\nIt runs as a command-line tool controlling a browser and reaching the network, so it declares shell execution, browser control, and outbound network — a combination that scores high in this catalog. CAPTCHA solving is worth naming plainly: it exists to defeat a control the site operator deliberately put in place, and building it into an agent stack is a choice about what that agent is for.", "website": "https://github.com/browser-act", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "unknown", "capabilities": [ "shell:execute", "browser:control", "network:outbound" ], "packaging": null, "configFiles": null, "category": "browser-automation", "updatedAt": "2026-09-02", "risk": { "score": 95, "band": "critical", "signals": [ { "id": "trust-unknown", "kind": "trust", "label": "Unknown publisher", "contribution": 20 }, { "id": "high-risk-capabilities", "kind": "capability", "label": "1 high-risk capability: shell:execute", "contribution": 10 }, { "id": "exfiltration-risk", "kind": "combination", "label": "Data exfiltration risk (high)", "description": "Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.", "contribution": 15 } ], "dangerousCombos": [ { "id": "exfiltration-risk", "name": "Data exfiltration risk", "description": "Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.", "capabilities": [ "shell:execute", "network:outbound" ], "severity": "high" } ] }, "dangerousCombos": [ { "id": "exfiltration-risk", "name": "Data exfiltration risk", "description": "Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.", "capabilities": [ "shell:execute", "network:outbound" ], "severity": "high" } ] }, { "kind": "skill", "slug": "anti-detect-browser", "name": "Anti-Detect Browser", "vendor": "antibrow", "description": "Launches browsers with distinct fingerprints so multiple accounts can be operated without the sites involved correlating them — built for multi-account automation.\n\nIt declares browser control and outbound network. This is the entry in the catalog where the risk is mostly about you rather than to you: evading fingerprinting is typically a terms-of-service violation on the target platform, and automating multiple accounts is prohibited outright by many. The technical capability is ordinary browser control; the exposure is account bans and, depending on jurisdiction and use, worse. Catalogued so that its presence in a stack is a deliberate decision.", "website": "https://github.com/antibrow", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "unknown", "capabilities": [ "browser:control", "network:outbound" ], "packaging": null, "configFiles": null, "category": "browser-automation", "updatedAt": "2026-09-02", "risk": { "score": 70, "band": "high", "signals": [ { "id": "trust-unknown", "kind": "trust", "label": "Unknown publisher", "contribution": 20 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "using-git-worktrees", "name": "Using Git Worktrees", "vendor": "obra/superpowers", "description": "Using Git Worktrees is a workspace-management skill from the Superpowers plugin that creates isolated git worktrees for feature work, keeping each branch in its own directory rather than constantly stashing and switching in a single checkout. It is intended for any task that needs isolation from the current workspace state — running a long-form refactor while continuing to ship hotfixes from main, executing an implementation plan in parallel with other work, or letting the agent operate without disturbing the user's local edits.\n\nFunctionally, the skill reads the current repository state to determine where to base the new worktree, writes a new worktree directory plus tracking metadata, and shells out to git through the Claude Code Bash tool to create, list, and clean up worktrees. It does not request network or credential access.\n\nThe skill is community trust and uses three declared capabilities — filesystem read, filesystem write, and shell execute. Shell execution is the one that lights up the larger risk signal, but because the shell calls are scoped to git operations and the skill is part of a well-known plugin, the practical risk is contained. It does match one half of the data-exfiltration dangerous combination on its own; the combination only activates when the running agent also has outbound network access.", "website": null, "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "filesystem:read", "filesystem:write", "shell:execute" ], "packaging": null, "configFiles": null, "category": null, "updatedAt": "2026-09-02T18:21:41.256Z", "risk": { "score": 65, "band": "high", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 }, { "id": "high-risk-capabilities", "kind": "capability", "label": "2 high-risk capabilities: filesystem:write, shell:execute", "contribution": 20 }, { "id": "filesystem-write-shell-risk", "kind": "combination", "label": "Persistence + execution risk (medium)", "description": "Shell execution plus filesystem write means the agent can plant persistent backdoors (e.g. modifying startup scripts).", "contribution": 5 } ], "dangerousCombos": [ { "id": "filesystem-write-shell-risk", "name": "Persistence + execution risk", "description": "Shell execution plus filesystem write means the agent can plant persistent backdoors (e.g. modifying startup scripts).", "capabilities": [ "shell:execute", "filesystem:write" ], "severity": "medium" } ] }, "dangerousCombos": [ { "id": "filesystem-write-shell-risk", "name": "Persistence + execution risk", "description": "Shell execution plus filesystem write means the agent can plant persistent backdoors (e.g. modifying startup scripts).", "capabilities": [ "shell:execute", "filesystem:write" ], "severity": "medium" } ] }, { "kind": "skill", "slug": "web-best-practices", "name": "Best Practices", "vendor": "addyosmani", "description": "Applies modern web development practices across security, performance, and code quality — the accumulated \"you should probably not do that\" knowledge that rarely makes it into a linter rule.\n\nIt reads and writes project files. Its usefulness depends on how current the encoded guidance is: web platform advice ages quickly, and a skill recommending a pattern that was correct three years ago is worse than no skill, because it arrives with apparent authority. Check the last update date before relying on it for security guidance specifically.", "website": "https://github.com/addyosmani", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "filesystem:read", "filesystem:write" ], "packaging": null, "configFiles": null, "category": "development", "updatedAt": "2026-09-02", "risk": { "score": 50, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 }, { "id": "high-risk-capabilities", "kind": "capability", "label": "1 high-risk capability: filesystem:write", "contribution": 10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "check-review", "name": "Check", "vendor": "tw93", "description": "Reviews code diffs, pull requests, and overall release readiness — pulling the change under review and reporting what looks risky before it ships.\n\nIt reads the repository and runs git and build commands, so it declares filesystem read and shell execution. Review skills are among the highest-value low-risk uses of an agent, because the output is advisory: a wrong finding costs a minute of human attention, where a wrong edit costs a debugging session. Keep it advisory — resist the version that fixes what it finds.", "website": "https://github.com/tw93", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "filesystem:read", "shell:execute" ], "packaging": null, "configFiles": null, "category": "code-review", "updatedAt": "2026-09-02", "risk": { "score": 50, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 }, { "id": "high-risk-capabilities", "kind": "capability", "label": "1 high-risk capability: shell:execute", "contribution": 10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "test-driven-development", "name": "Test-Driven Development", "vendor": "obra/superpowers", "description": "Test-Driven Development is a rigid process skill from the Superpowers plugin that enforces a strict red-green-refactor loop for any feature implementation or bug fix. Before the agent is allowed to write production code, the skill requires it to write a failing test that captures the desired behavior, watch that test fail for the right reason, write the minimum implementation that makes the test pass, and only then refactor with the test still passing as a guardrail.\n\nFunctionally, the skill reads existing test and source files to find the right place for new tests and code, and writes new test files plus implementation files as the agent moves through the loop. It does not request shell execution — the test runs themselves go through the Claude Code Bash tool — and it does not request network or credential access.\n\nThe skill is community trust because it ships in an independent plugin. The active risk signals come from filesystem read and write, the minimum surface for any skill that produces tests and code, and do not match any dangerous combination. The notable property of the skill is not its capability surface but its discipline: it actively prevents agents from skipping the test step when the underlying task is harder than expected.", "website": null, "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "filesystem:read", "filesystem:write" ], "packaging": null, "configFiles": null, "category": null, "updatedAt": "2026-09-02T18:21:41.256Z", "risk": { "score": 50, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 }, { "id": "high-risk-capabilities", "kind": "capability", "label": "1 high-risk capability: filesystem:write", "contribution": 10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "writing-skills", "name": "Writing Skills", "vendor": "obra/superpowers", "description": "Writing Skills is the meta-skill from the Superpowers plugin: it is the skill the agent uses when the task is to create a new skill, edit an existing skill, or verify a skill works before it ships. It encodes the structure that every other Superpowers skill follows — frontmatter, description, trigger conditions, body content, and any associated reference material — and provides the verification steps that confirm a new skill is wired into the plugin correctly.\n\nFunctionally, the skill reads existing skills as examples and templates and writes new or edited skill files into the plugin's skill directory. It does not request shell execution, network access, or credentials, and it does not invoke other skills — it produces them.\n\nThe skill is community trust and declares only filesystem read and write capabilities, which produce a routine risk signal and do not match any dangerous combination. Its practical importance is structural: because every other Superpowers skill is created through this skill, it is the right place to validate conventions, enforce naming, and catch trigger-condition mistakes that would otherwise cause a new skill to silently never activate.", "website": null, "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "filesystem:read", "filesystem:write" ], "packaging": null, "configFiles": null, "category": null, "updatedAt": "2026-09-02T18:21:41.256Z", "risk": { "score": 50, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 }, { "id": "high-risk-capabilities", "kind": "capability", "label": "1 high-risk capability: filesystem:write", "contribution": 10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "azure-deploy", "name": "Azure Deploy", "vendor": "Microsoft", "description": "Microsoft's skill for executing Azure deployments with validation and error recovery — provisioning resources, handling partial failures, and retrying with corrected parameters rather than stopping at the first error.\n\nIt runs deployment commands, holds cloud credentials, and reaches the network, so it declares shell execution, credential access, and outbound network. Shell plus network is Anomity's exfiltration pattern, and the credentials in play are cloud-plane credentials — which is to say this is one of the highest-consequence skills in the catalog, and the one most deserving of a human approving each run rather than a standing grant.", "website": "https://github.com/microsoft", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "official", "capabilities": [ "shell:execute", "credentials:access", "network:outbound" ], "packaging": null, "configFiles": null, "category": "devops", "updatedAt": "2026-09-02", "risk": { "score": 45, "band": "medium", "signals": [ { "id": "trust-official", "kind": "trust", "label": "Published by the vendor", "contribution": -40 }, { "id": "high-risk-capabilities", "kind": "capability", "label": "2 high-risk capabilities: shell:execute, credentials:access", "contribution": 20 }, { "id": "exfiltration-risk", "kind": "combination", "label": "Data exfiltration risk (high)", "description": "Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.", "contribution": 15 } ], "dangerousCombos": [ { "id": "exfiltration-risk", "name": "Data exfiltration risk", "description": "Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.", "capabilities": [ "shell:execute", "network:outbound" ], "severity": "high" } ] }, "dangerousCombos": [ { "id": "exfiltration-risk", "name": "Data exfiltration risk", "description": "Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.", "capabilities": [ "shell:execute", "network:outbound" ], "severity": "high" } ] }, { "kind": "skill", "slug": "azure-diagnostics", "name": "Azure Diagnostics", "vendor": "Microsoft", "description": "Diagnoses and remediates Azure production issues systematically — reading resource health, correlating logs, and proposing or applying fixes for common failure modes.\n\nCapabilities are shell execution, credential access, and outbound network. The remediation half is what makes it more than a reporting tool and more than a read-only risk: a skill that can change production configuration during an incident is operating at the moment when review is least likely to happen. Prefer running it in diagnose-only mode and letting a human apply the change.", "website": "https://github.com/microsoft", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "official", "capabilities": [ "shell:execute", "credentials:access", "network:outbound" ], "packaging": null, "configFiles": null, "category": "devops", "updatedAt": "2026-09-02", "risk": { "score": 45, "band": "medium", "signals": [ { "id": "trust-official", "kind": "trust", "label": "Published by the vendor", "contribution": -40 }, { "id": "high-risk-capabilities", "kind": "capability", "label": "2 high-risk capabilities: shell:execute, credentials:access", "contribution": 20 }, { "id": "exfiltration-risk", "kind": "combination", "label": "Data exfiltration risk (high)", "description": "Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.", "contribution": 15 } ], "dangerousCombos": [ { "id": "exfiltration-risk", "name": "Data exfiltration risk", "description": "Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.", "capabilities": [ "shell:execute", "network:outbound" ], "severity": "high" } ] }, "dangerousCombos": [ { "id": "exfiltration-risk", "name": "Data exfiltration risk", "description": "Shell execution combined with outbound network access can exfiltrate arbitrary data from the machine.", "capabilities": [ "shell:execute", "network:outbound" ], "severity": "high" } ] }, { "kind": "skill", "slug": "ab-testing", "name": "A/B Testing", "vendor": "coreyhaines31", "description": "Plans and designs experiments — hypothesis, metric selection, sample size, and the analysis plan — before anything is shipped, which is where most experimentation programmes go wrong.\n\nIt declares no capabilities; it produces a plan rather than touching systems, so it adds no risk signal to the running agent. Its value is in the discipline it imposes: an agent that will happily invent a plausible-sounding lift number is much less dangerous when it has been made to commit to a metric and a sample size up front.", "website": "https://github.com/coreyhaines31", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [], "packaging": null, "configFiles": null, "category": "data-analysis", "updatedAt": "2026-09-02", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "accessibility-audit", "name": "Accessibility", "vendor": "addyosmani", "description": "Audits and improves web accessibility against WCAG 2.2, checking contrast, focus order, labelling, and semantics, then proposing concrete fixes rather than a list of violations.\n\nIt drives a browser to evaluate rendered pages and reads project source to suggest changes, so it declares browser control and filesystem read. Accessibility work is one of the better fits for an agent — the rules are objective, the fixes are mechanical, and a human still has to judge whether the result is genuinely usable rather than merely conformant.", "website": "https://github.com/addyosmani", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "browser:control", "filesystem:read" ], "packaging": null, "configFiles": null, "category": "testing", "updatedAt": "2026-09-02", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "agent-pulse", "name": "Agent Pulse", "vendor": "jane-o-o-o-o", "description": "Inspects local AI-agent activity logs across multiple platforms, reporting sessions, token usage, and estimated cost — visibility into what the agents on a machine have actually been doing.\n\nIt reads local log and configuration files, so it declares filesystem read. It is catalogued partly as a useful tool and partly as a reminder: agent logs are a rich forensic source containing prompts, file paths, and sometimes the contents of what was worked on. That makes them valuable for incident review and sensitive if they leave the machine.", "website": "https://github.com/jane-o-o-o-o", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "filesystem:read" ], "packaging": null, "configFiles": null, "category": "observability", "updatedAt": "2026-09-02", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "ai-seo", "name": "AI SEO", "vendor": "coreyhaines31", "description": "Optimises content for AI search engines and LLM citation — structuring pages so a model can extract a clean answer and attribute it, rather than optimising purely for classical ranking signals.\n\nOutbound network access is its only declared capability, for fetching and checking pages. Worth a caveat this catalog is well placed to give: much published advice in this area is ahead of the evidence. Structured, concise, well-sourced content genuinely helps extraction; several widely-promoted mechanisms have no demonstrated effect on citation at all.", "website": "https://github.com/coreyhaines31", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "network:outbound" ], "packaging": null, "configFiles": null, "category": "marketing", "updatedAt": "2026-09-02", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "ask-matt", "name": "Ask Matt", "vendor": "mattpocock", "description": "A router skill: it reads the user's situation and points them at the skill that actually fits, instead of every skill competing to be invoked for every prompt.\n\nIt declares no capabilities — it only decides what to call next — which makes it one of the zero-risk entries in the catalog. The pattern is more interesting than the skill: as skill libraries grow into the hundreds, selection becomes the bottleneck, and an explicit router is a more legible answer than hoping description-matching picks correctly.", "website": "https://github.com/mattpocock", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [], "packaging": null, "configFiles": null, "category": "productivity", "updatedAt": "2026-09-02", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "audit-website", "name": "Audit Website", "vendor": "squirrelscan", "description": "Runs a large rule set — reportedly 249+ checks — over a site covering SEO, performance, security headers, and accessibility, then drives iterative fixes rather than only reporting.\n\nIt fetches and renders pages, so it declares browser control and outbound network access. Breadth is both the appeal and the caveat: a suite this wide will produce findings that are technically true and practically irrelevant, and an agent that fixes everything it is told about can churn a codebase for little benefit. Triage the report before authorising the fixes.", "website": "https://squirrelscan.com", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "browser:control", "network:outbound" ], "packaging": null, "configFiles": null, "category": "testing", "updatedAt": "2026-09-02", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "brainstorming", "name": "Brainstorming", "vendor": "obra/superpowers", "description": "Brainstorming is a process skill from the Superpowers plugin that runs before any creative or generative work — building a feature, designing a component, adding behavior, or modifying an existing flow. Instead of letting the agent dive straight into code from a short user prompt, the skill forces a deliberate exploration of intent, requirements, and design alternatives. It walks the user through clarifying the actual problem, naming the constraints, sketching at least two viable approaches, and recording the decisions before any file is touched.\n\nThe skill itself does not read or write files — it operates entirely in the conversation — but it shapes everything the agent does afterward by anchoring later steps to the decisions surfaced during the brainstorm. Because it has no declared filesystem, shell, or network capabilities of its own, it contributes no extra risk signals to the running agent. It is one of the safest skills in the catalog and one of the highest-leverage ones for avoiding the half-finished, scope-creeping implementations that result when an agent jumps from a vague prompt straight into code.", "website": null, "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [], "packaging": null, "configFiles": null, "category": null, "updatedAt": "2026-09-02T18:21:41.256Z", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "browser-testing-with-devtools", "name": "Browser Testing with DevTools", "vendor": "addyosmani", "description": "Tests in a real browser via Chrome DevTools — inspecting the DOM, watching network activity, and reproducing behaviour the way a developer would rather than through a synthetic test harness.\n\nDeclared capability is browser control. The distinction from a headless test runner matters for risk: a skill driving your actual browser can reach internal hosts and act inside authenticated sessions. Point it at a dedicated profile. As a debugging aid it is genuinely strong, because it observes what the page does instead of what the test asserts.", "website": "https://github.com/addyosmani", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "browser:control" ], "packaging": null, "configFiles": null, "category": "testing", "updatedAt": "2026-09-02", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "caveman", "name": "Caveman", "vendor": "juliusbrussee", "description": "An ultra-compressed communication mode that strips agent output to essentials, reportedly cutting token usage by around 75% by dropping hedging, restatement, and prose scaffolding.\n\nIt declares no capabilities — it changes how the agent writes, not what it touches. The trade is legibility for cost: compressed output is cheaper and faster to skim, and also easier to misread, because the qualifiers that carried the uncertainty are the first thing removed. Reasonable for routine work, less so for anything where the caveats mattered.", "website": "https://github.com/juliusbrussee", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [], "packaging": null, "configFiles": null, "category": "productivity", "updatedAt": "2026-09-02", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "caveman-commit", "name": "Caveman Commit", "vendor": "juliusbrussee", "description": "Generates compressed commit messages in Conventional Commits format from the staged diff, aiming for a parseable subject line without the paragraph of narration agents tend to produce.\n\nIt reads the repository and works through git, so it declares filesystem read and git access. Machine-readable commit messages are genuinely useful downstream for changelogs and release automation. The caution is that a message generated from a diff describes what changed, not why — and the why is the part a future reader actually needs.", "website": "https://github.com/juliusbrussee", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "filesystem:read", "git:access" ], "packaging": null, "configFiles": null, "category": "development", "updatedAt": "2026-09-02", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "deep-research", "name": "Deep Research", "vendor": "samber", "description": "Conducts parallel web searches with multi-source validation and produces a cited report, deliberately cross-checking claims across sources rather than summarising the first result.\n\nOutbound network access is its only declared capability. Multi-source validation is a real defence against both hallucination and retrieval-borne prompt injection — a claim that appears in one adversarial page and nowhere else fails corroboration. It is not a complete defence, since sources can agree by copying each other, but requiring citations at least makes the failure inspectable.", "website": "https://github.com/samber", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "network:outbound" ], "packaging": null, "configFiles": null, "category": "research", "updatedAt": "2026-09-02", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "notebooklm-skill", "name": "NotebookLM Skill", "vendor": "PleasePrompto", "description": "Lets an agent query NotebookLM for source-grounded answers over documents already uploaded there, so responses are anchored to a specific corpus rather than to model recall.\n\nIt reaches the service over the network and drives the browser session that holds the notebook, so it declares browser control and outbound network. Source grounding is a real accuracy improvement and a real data-flow decision: the questions asked, and the notebook contents that answer them, move between two AI systems. Worth being deliberate about which documents live in that notebook.", "website": "https://notebooklm.google.com", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "browser:control", "network:outbound" ], "packaging": null, "configFiles": null, "category": "research", "updatedAt": "2026-09-02", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "systematic-debugging", "name": "Systematic Debugging", "vendor": "obra/superpowers", "description": "Systematic Debugging is a process skill from the Superpowers plugin that runs before the agent is allowed to propose any fix for a reported bug, test failure, or unexpected behavior. Rather than letting the agent jump to its first hypothesis, the skill walks through a deliberate sequence: reproduce the problem under controlled conditions, narrow the failing surface to the smallest input that still fails, capture the actual versus expected output, form a hypothesis, and only then change code.\n\nThe skill reads source files, test outputs, and any artifacts the user has shared to characterize the bug, but it does not write files or execute commands — it shapes the agent's reasoning rather than the code itself. It declares only a filesystem read capability, which produces a low risk signal and does not match any dangerous combination.\n\nThe value of the skill is in what it prevents: shipping a fix for the wrong cause, declaring a flaky test \"fixed\" without understanding why, and the long detours that happen when an agent commits to a hypothesis before it has actually reproduced the failure.", "website": null, "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "filesystem:read" ], "packaging": null, "configFiles": null, "category": null, "updatedAt": "2026-09-02T18:21:41.256Z", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] }, { "kind": "skill", "slug": "youtube-transcript-downloader", "name": "YouTube Transcript Downloader", "vendor": "michalparkola", "description": "Retrieves and processes transcripts from YouTube URLs so an agent can work with the contents of a talk or tutorial as text instead of being told to watch a video it cannot watch.\n\nOutbound network access is its only capability, making it a low-risk utility. The standard retrieval caveat applies and is easy to forget with video: a transcript is text an unknown person wrote, and an agent that summarises it is processing untrusted input. Video feels less like a document than a web page, which is precisely why it is a convenient injection surface.", "website": "https://github.com/michalparkola", "repository": null, "documentationUrl": null, "license": null, "version": null, "trustLevel": "community", "capabilities": [ "network:outbound" ], "packaging": null, "configFiles": null, "category": "research", "updatedAt": "2026-09-02", "risk": { "score": 40, "band": "medium", "signals": [ { "id": "trust-community", "kind": "trust", "label": "Established community project", "contribution": -10 } ], "dangerousCombos": [] }, "dangerousCombos": [] } ] }