Choose development skills by task, source and evidence
Choose development skills by the task you need to complete. A catalog category can narrow a search, but the original instructions, required tools and evidence for the current package determine whether a resource is a useful starting point.
Begin with a small, observable outcome
Write one sentence that describes the result you need. “Add a Windows settings screen that opens and responds to keyboard input” is more useful than “find the best coding skill.” So is “isolate a feature branch without disturbing my existing work” or “review the dependencies in a proposed CI workflow.” Each gives you a different selection criterion and a result you can check.
Record the environment beside that outcome: operating system, framework, agent client, repository state and any services the task can access. Decide whether this is a read-only review, a local change or a deployment. A resource that describes deployment may require account access and can publish a project; a planning resource may produce a specification without changing any application files.
Our Development collection, Web collection and Workflow collection group these tasks broadly. Use the library to narrow the use case, then inspect the selected resource. A category is an editorial navigation choice, not a claim that every resource in it performs the same job.
Separate planning, implementation and diagnostics
The Superpowers resources cover several stages of development. They are related, but they are not interchangeable. The following examples describe the purpose of the pinned versions in this catalog; they do not report a successful agent run.
| Your immediate task | Resource to inspect | What to check before using it |
|---|---|---|
| Clarify a proposed feature and its requirements | Brainstorming | What decisions and approval checkpoints it asks for |
| Find and invoke the relevant skills in that ecosystem | Using Superpowers | Which harness tools and other skills it expects |
| Isolate feature work from the current checkout | Using Git Worktrees | Existing isolation, repository state and platform-specific workspace tools |
| Execute an existing implementation plan in the current session | Executing Plans | Whether the plan is already approved and how completion will be reviewed |
| Investigate a problematic agent session | Diagnosing Superpowers | Transcript access and sensitive information in any exported report |
A useful selection process follows the stage you are actually in. Do not install every related resource because their names appear in the same repository. Read how they depend on one another and whether your client provides the expected tools. Some instructions are designed for a larger plugin or workflow; an individual download is not evidence that its complete ecosystem has been installed.
These distinctions are based on the pinned Superpowers source repository. We reviewed the relevant original SKILL.md files to correct their catalog categories. That review covers classification, not runtime behavior, transcript privacy or the safety of every instruction.
Match the application platform and deployment stage
A desktop interface and a website can both have buttons, forms and navigation, but they require different toolchains. WinUI App concerns Windows desktop development. Its pinned instructions describe C#, WinUI 3 and the Windows App SDK, including environment setup and application launch verification. It belongs in development browsing rather than a collection of website design resources.
ASP.NET Core covers development and review of web applications. Vercel Deploy addresses publishing applications and websites. Both are relevant to web work; neither is simply a UI styling guide. Choose between them by whether you need to build or review application behavior, or deploy an already prepared project.
For these examples, inspect the pinned OpenAI skill sources. Read the resource's tool requirements and any account or network access it requests. A free download does not make third-party hosting, model usage or other services free. Set those cost and access boundaries before running commands.
Treat client labels as browsing metadata
The library's Catalog client filter matches a complete label, such as Codex or Claude Code. It does not interpret a partial word as another client name. You can combine the filter with a task, author, license or output collection, and the same filters apply to the public catalog API.
Client labels help you discover candidates. They are catalog metadata, not compatibility test results. A client may recognize a SKILL.md file while lacking the shell, browser, plugin hooks or named tools used by its instructions. Check the upstream installation path and prerequisites for your actual client rather than assuming that a matching label proves an end-to-end integration.
Search can also find an exact resource identifier such as obra-using-git-worktrees. Percent signs, underscores and backslashes are treated as literal search characters. If a combination returns no results, remove one filter or search a more general task phrase. A blank result should not be interpreted as proof that a tool cannot perform the task.
Compare source, package and current evidence
Use Compare resources to put up to three candidates beside each other. Check their purpose, source, license, version, package inventory and prerequisites. Open the original source and the packaged instruction files before deciding what to install.
A pinned source identifies the revision that was used. A SHA-256 checksum helps establish whether the package bytes have changed. A license describes permitted use under its terms. These are different facts: none proves that the instructions achieve your goal or that all external actions are safe.
Read any scenario result together with its environment, version and checksum. A local static check can establish a narrow property of a fixture without exercising a whole skill through an AI client. For example, the GitHub Actions Efficiency Review has a bounded local workflow-lint record. Its result does not measure real cache savings or certify every workflow. The CI review guide explains that boundary in more detail.
An absent result means that no published outcome record matches the current package in this catalog. It is not an automatic failure. A pass is equally not a guarantee for another repository, client or environment. If you need a specific behavior, look for evidence of that behavior rather than a badge or a total count.
Try one candidate with a reversible fixture
Prepare a small copy of the work or an isolated test repository. Remove secrets and production access from the fixture. For a read-only task, decide what the tool may inspect. For a change, define the allowed files and retain a clean baseline. For deployment, identify the target and review the actual change before publishing.
Write an expected result before the run: a launchable local window, an unchanged primary checkout, a reviewed plan, or a diagnostic report with specific supporting evidence. Afterward, inspect the changed files and compare the output with that expectation. Include a negative case when it is useful, such as an unsupported environment or a missing prerequisite, so you can see whether the workflow stops clearly.
Record what you ran, the client and tool versions, the package version and checksum, and the limits of the observation. Keep personal information and secrets out of public evidence. Start with one candidate; add another only when it addresses a concrete missing step. This makes a comparison useful without turning a large catalog into a list of unexamined installations.
What this guide establishes
This is an original BB Skills selection guide, written with AI assistance and checked against the pinned source files and the site's filter behavior. The examples explain catalog purposes and evaluation steps. No AI client execution, Windows application build, Superpowers session analysis or Vercel deployment was performed for this guide.
The classification corrections preserve the existing resource identifiers, package bytes, versions, licenses and recorded outcomes. Their purpose is to make navigation reflect the work described in the instructions. Revisit the source and your own environment when choosing a resource; this guide is a starting point for that review.