Claude Code has become one of the most closely watched AI coding tools because it does more than autocomplete code. It can read a repository, edit files, run commands, and work with project context in a way that feels much closer to an operational coding agent than a traditional assistant. But the more teams use it in real workflows, the clearer one point becomes: the model is only part of the story. A lot of the real leverage now comes from the structure built around it. Anthropic’s own documentation shows that Claude Code is no longer just about CLAUDE.md, but also about settings, hooks, skills, commands, subagents, rules, and project memory inside .claude/ and ~/.claude/.
That shift matters because many public discussions still reduce Claude Code to a prompt file and a few handwritten instructions. In practice, the official product already supports a much richer operating layer. Anthropic documents native skills, reusable project-level configuration, and subagents with isolated task scopes, while hooks can enforce behavior at specific moments instead of relying on the model to remember and follow every instruction probabilistically. That is exactly why community repositories are attracting so much attention: they are not changing the model itself, but they are changing how it works day to day.
Get Shit Done and the push to fix context decay
One of the best-known examples is Get Shit Done, usually shortened to GSD. The project describes itself as a lightweight but powerful meta-prompting, context-engineering, and spec-driven development system for Claude Code and other agentic coding tools. Its core claim is simple and compelling: it tries to solve “context rot,” the gradual drop in quality that happens as long sessions fill up Claude’s context window. The repository frames the problem directly and presents its answer as a combination of context engineering, XML prompt formatting, subagent orchestration, and state management.
What makes GSD interesting is not that it promises magic, but that it tries to impose discipline on long-running work. Instead of letting one oversized conversation carry everything, it breaks work into smaller plans, externalizes state into files, and gives Claude a more structured way to continue from clean or cleaner context. For developers who have already noticed that a session can begin sharply and then become less reliable as it grows, that approach makes immediate sense. The repo also supports a one-command install path through npx get-shit-done-cc, which has helped it spread quickly among Claude Code users.
There is a practical caveat, though. GSD’s workflow is opinionated, and its benefits depend on whether a team is willing to adopt that structure rather than just borrowing one or two ideas from it. It is best understood not as a single prompt pack, but as a process framework for getting more predictable output from Claude Code on complex tasks. That distinction matters, because the strongest Claude setups increasingly look less like clever prompting and more like lightweight operating systems for an agent.
Antigravity Awesome Skills and the rise of installable playbooks
The second project worth watching is Antigravity Awesome Skills, which positions itself as an installable library of reusable SKILL.md playbooks for Claude Code and other AI coding assistants. Its current README describes a catalog of more than 1.377 skills, plus bundles, workflows, installer paths, and plugin-safe distributions. It also provides a direct Claude Code installation path through npx antigravity-awesome-skills --claude.
This matters because skills are no longer a community workaround; they are an official Claude Code feature. Anthropic documents skills as a native way to extend Claude’s capabilities and states that custom commands have effectively been merged into the same system. A file in .claude/commands/ and a skill in .claude/skills/.../SKILL.md can both create slash-style commands and behave the same way. That means repositories like Antigravity are building on top of a supported product primitive, not inventing a parallel ecosystem from scratch.
The appeal is obvious. Instead of collecting disconnected prompt snippets in notes or bookmarks, teams can install reusable playbooks for planning, coding, debugging, security review, infrastructure work, product tasks, or growth operations. In that sense, Antigravity works like a packaging layer for repeatable agent behavior. The risk, of course, is the same one that affects every large toolkit: abundance can become noise. A library with more than a thousand skills is powerful, but only if developers know which pieces genuinely fit their workflow and which ones just add clutter.
Awesome Claude Code Toolkit as a map of the ecosystem
The third repository, Awesome Claude Code Toolkit, takes a broader approach. Rather than focusing only on one framework or one installable library, it tries to serve as a more comprehensive Claude Code toolkit. The project describes itself as a collection that includes 135 agents, 35 curated skills, 42 commands, 150+ plugins, 19 hooks, 15 rules, templates, MCP configurations, and more.
That breadth is what makes it valuable. It reflects how Claude Code is being used by power users now: not as a single assistant with one memory file, but as a configurable environment made up of roles, reusable operating instructions, lifecycle hooks, and project-specific conventions. Anthropic’s own docs support that framing. The .claude directory is where Claude Code reads instructions, settings, hooks, skills, commands, subagents, rules, and memory, while hooks are explicitly documented as a way to automate actions that should happen every time, without exception.
That is probably the biggest lesson across all three repositories. Claude Code gets more useful when fewer important behaviors are left to chance. GSD tries to make long tasks more reliable by controlling context. Antigravity packages reusable skills into installable building blocks. Awesome Claude Code Toolkit offers a broader map of agents, hooks, plugins, and conventions that can help teams assemble a more serious setup. None of these repos replaces engineering judgment, but together they show where the ecosystem is heading: away from one-off prompting and toward structured, shareable, operational systems for AI-assisted software development.
FAQ
Is Claude Code still mainly driven by CLAUDE.md?CLAUDE.md is still important, but Anthropic documents a much broader structure. Claude Code also reads settings, hooks, skills, commands, subagents, rules, and memory from .claude/ and ~/.claude/.
What does Get Shit Done add to Claude Code?
According to its repository, GSD is designed to reduce context decay by combining meta-prompting, context engineering, spec-driven development, subagent orchestration, and externalized state management. It is aimed at making long or complex workflows more reliable.
Are Claude Code skills an official feature or just a community hack?
They are official. Anthropic documents skills as a native Claude Code feature and notes that custom commands have been folded into the same system, so skills and command-style extensions can work in equivalent ways.
Why are these repositories useful for teams, not just solo developers?
Because Claude Code supports project-level configuration that can be committed and shared. Anthropic notes that project files inside .claude/ can be versioned in git so teams can share skills, settings, and other behavior across a repository.