# my-skills: a personal marketplace of Claude Code skills for scaffolding, review and documentation

> bear2u/my-skills is a collection of Claude Code custom skills published as a plugin marketplace, covering project scaffolding for Flutter and Next.js, a Codex driven review loop, and change documentation served in the browser. It is a personal toolbox rather than a library, which is both its appeal and its main caveat.

**bear2u/my-skills** — 자주 사용하는 스킬들을 모은 나의 클로드 스킬들입니다. 

- Repository: https://github.com/bear2u/my-skills
- Stars: 934 · Forks: 209
- Language: HTML
- License: not declared
- Published: 2026-10-20 · Updated: 2026-10-20 · Language: en
- Canonical page: https://hysenlabs.com/projects/bear2u-my-skills

## A personal toolbox published as a plugin marketplace, not a library

The project description says it plainly, in Korean: a collection of the author's own Claude Code skills, gathered because they are used often. That framing sets expectations correctly. There is no package on npm, no published type definitions, and no API. What exists is a repository of skill folders under a `skills/` directory, each with its own directory entry, plus a plugin marketplace that lets Claude Code install them by name.

The README states the purpose in one line: custom skills for Claude Code that provide various automation skills aimed at raising development productivity. The audience is therefore a single developer or a small team already inside Claude Code, who keeps writing the same scaffolding prompts by hand and wants them packaged once.

The repository layout supports that reading. Alongside `skills/` and `.claude/` there are folders that look like working notes rather than released artifacts: `backup/`, `chatgpt-web-research/`, `claude-3.5-sonnet`, `landing-page-v2-test/`, `test-project/`, `workthrougt-test/`, plus loose files such as `hello.txt`, `Helloworld.txt`, `extract-markdown.js`, `fetch-dynamic-content.js` and `flutter-animation-guide.md`. There is a `SKILL_DISTRIBUTION_GUIDE.md` describing how the skills are distributed. `node_modules/` and `package-lock.json` are committed too. This is a working repository that happens to be the distribution channel, which is the central fact to keep in mind when judging it.

## Adding the marketplace and installing a skill by slug

Installation uses the plugin marketplace commands of Claude Code. The README gives the marketplace add step in a single bash block:

```bash
/plugin marketplace add bear2u/claude-plugins
```

After the marketplace is registered, each skill installs with its own command. The README lists nineteen of them in a table, each row giving a one line description and the exact install command. A representative slice:

```bash
/plugin marketplace install suji-father-marketplace@flutter-init
/plugin marketplace install suji-father-marketplace@code-changelog
/plugin marketplace install suji-father-marketplace@codex
```

Notice that the marketplace identifier inside the command is `suji-father-marketplace`, not the repository owner, even though the marketplace is added from `bear2u/claude-plugins`. The two names do not match, which is a small trap for anyone scripting the install, and it is also a hint about how these skills are organised: the marketplace is the distribution point and each skill is a separate entry inside it.

The rest of the table shows how broad the collection is. There are project generators for Flutter and Next.js, prompt tooling (`meta-prompt-generator`, `prompt-enhancer`, `code-prompt-coach`), card news image generators for Instagram, a landing page design guide, a workthrough skill that documents work and, in its second version, serves the documentation with VitePress on port 5173, a DuckDuckGo based search skill with region and period filters, and a web page to Markdown converter with normal, AI optimised and dual modes. There are also skills that hand work to the OpenAI Codex CLI.

## Code Changelog keeps a Markdown record and serves it in the browser

Code Changelog is the clearest single mechanism in the collection, and worth reading the description of closely because it shows the pattern several other skills follow. It documents all AI generated code changes automatically and lets you review them live in a web browser.

The mechanism described is a two part output. The first is automatic Markdown generation, so the record of what changed is plain text on disk and stays diffable and greppable. The second is an HTML viewer that needs no installation and runs on a Python server, with a dark mode interface in GitHub style, listening on `http://localhost:4000`.

The stated use cases are code review documentation, change history tracking, and sharing changes with a teammate. The realtime server is the piece that matters for review work, since it means a reviewer can watch documentation accumulate during the session instead of waiting for a commit. The choice of a Python server for the viewer, with no build step, keeps the dependency surface of the skill small, which fits a tool meant to sit in the background of an unrelated coding task.

This is also the clearest example of a family resemblance across the collection. The workthrough skill in its second version does something structurally similar: it documents development and modification work and then serves it as a VitePress build on port 5173. Both turn an agent session into a browsable artefact, one with a Python server and plain HTML, the other with a VitePress site.

## Codex as the verifier in a two model loop, with three sandbox levels

The Codex skill is a bridge to the OpenAI Codex CLI. According to the README it handles code analysis, refactoring and automated editing, and it exposes the CLI surface a power user needs: selection of the conversational model and reasoning level (`gpt-5`, `gpt-5-codex`), session resumption through `codex exec resume --last`, and automated editing through `--full-auto`.

The sandbox modes are the part with real consequences. The README lists three. `read-only` is for analysis alone and only reads. `workspace-write` allows modifying local files. `danger-full-access` grants full permissions including network access. A skill that flips a CLI into the third mode deserves a deliberate choice rather than a default, and the naming makes the gradient obvious.

Codex-Claude Loop builds on this by assigning roles: Claude plans and writes all code, Codex verifies. The documented cycle is Claude planning and implementing, then Codex validating, then feedback, then Claude revising, then Codex re-verifying, repeating. The verification pass is described as checking the plan for logic errors and security weaknesses, and the implementation for bugs, performance and security. The README also says plainly when not to use it: simple one-off edits and prototype code are too much for it.

The shell lines that drive the loop are given in the README:

```bash
echo "Review this plan..." | codex exec -m gpt-5-codex --config model_reasoning_effort="high" --sandbox read-only
echo "Review implementation..." | codex exec --sandbox read-only
echo "Verify fixes..." | codex exec resume --last
```

Every Codex invocation in that loop runs with `--sandbox read-only`. The dangerous mode is available in the base skill but is not part of the loop shown here.

## Scaffolding skills bake in specific stacks, not just file trees

The two project generators are where the collection stops being a set of prompts and becomes a set of opinions. The Flutter Init skill generates a domain based Flutter project using Clean Architecture, and the README lists what that means concretely: domain choice (Todo, Habit, Note, Expense or Custom), a stack preset (Minimal, Essential, Full Stack, Custom), and a wired stack of Riverpod 3.x for state, Drift for the local database, Freezed for immutable models, Easy Localization for languages and FluentUI icons.

Next.js 15 Init does the parallel thing with the App Router and TypeScript strict mode, offering domains of Todo, Blog, Dashboard or E-commerce, and pulling in ShadCN/ui for components, Zustand for client state, Tanstack Query for server state, Drizzle ORM and Better Auth for authentication.

Both are opinionated by design, which is the point and also the limit. The preset lists tell you the default assumptions: a Korean friendly localisation layer, a specific state library, a specific ORM. If your team has already chosen differently, the generator's value drops sharply, because you end up unpicking its choices rather than inheriting them.

The Prompt Enhancer skill sits upstream of both. It analyses the project context and turns a short request into a detailed specification, with the README giving a plain example: a request like "make me a login feature" becomes a full implementation requirement, including a Clean Architecture proposal and the application of project conventions. Its supported stacks are Flutter with Clean Architecture and Riverpod, Next.js and React with App Router and Zustand, and Python with Django and FastAPI, so it is not a general purpose request expander.

## Web facing skills rely on public endpoints and a script side project

Two skills reach outside the repository. Web Search is described as a DuckDuckGo based text, news and image search with region and period filters. DuckDuckGo is the whole backend, so there is no index of its own to maintain and no search key to configure, but there is equally no guarantee about result quality or rate limits, since the README documents neither.

Web to Markdown converts a web page into Markdown and offers three modes: normal, AI optimised and dual. The dual mode is the interesting one, since it suggests two conversions of the same page, one for a human reader and one prepared for a model to consume.

The repository itself contains the JavaScript that supports this kind of work. `extract-markdown.js` and `fetch-dynamic-content.js` sit at the top level next to `flutter-animation-guide.md` and `codedeck-flutter-animation.html`, and there is a `compose` directory and an `imgs/` folder for assets. The `chatgpt-web-research/` directory and the `claude-3.5-sonnet` folder point to experiments with other model interfaces sitting alongside the skills themselves.

The documentation work is one more of these side paths. The workthrough skill stores work in a workthrough format, and its second version adds a VitePress build served on port 5173, so a session can be turned into browsable documentation without hand writing a site.

## No releases, no populated licence, and a commit history that stops in June

The maintenance picture is thin, and honesty about it matters more than a favourable reading. The repository has no GitHub releases at all, so there is no version history to pin against and no upgrade notes. The last push was on 2026-06-22. The default branch is `master`, and the licence field on the repository is not populated, which leaves the terms of reuse genuinely unclear: you can read the code and follow the install commands, but nothing in the repository states a licence you can rely on, and that question needs an answer from the author before any skill body is copied into a commercial codebase.

The project also spans several audiences at once, and the README reflects that. A card news generator for Instagram, a landing page design guide, a Midjourney prompt generator and a Flutter animation guide are creative and content oriented, while Clean Architecture scaffolds, a Codex review loop and a changelog viewer are engineering tasks. The SEO and content skills are aimed at producing web pages and images, the code skills at producing and checking source files, and the prompt skills at improving how you talk to the model in the first place.

Given that mix, the practical question is not whether the collection is good but which two or three entries you actually need. Installing all nineteen from the marketplace would put far more prompt surface in every session than most developers want.

## Conclusion

my-skills suits a developer working in Claude Code who wants ready made scaffolding and a second opinion loop without writing each skill by hand, and it is the wrong choice for a team that needs versioned, reviewed, licensed tooling, since the repository has no releases, the licence field is not populated, and the skill code sits under a personal marketplace name rather than a maintained package. Check three things before adopting anything from it: that the marketplace add command succeeds for your Claude Code setup, that the skill you want matches your stack (Riverpod and Drizzle for the Next.js and Flutter scaffolds, DuckDuckGo for the search skill), and that the licence field on the repository is empty before you copy any skill body into a commercial codebase.

## FAQ

### How do I install my-skills into Claude Code?

Add the marketplace first with `/plugin marketplace add bear2u/claude-plugins`, then install individual skills with commands such as `/plugin marketplace install suji-father-marketplace@flutter-init`. The marketplace identifier in the install command is suji-father-marketplace even though the marketplace is added from the bear2u/claude-plugins repository.

### Does my-skills support anything other than Claude Code?

The skills are written for Claude Code, and one of them, Codex-Claude Loop, calls the OpenAI Codex CLI from inside a Claude session. The repository also carries a chatgpt-web-research folder, but the README describes the collection as Claude Code custom skills.

### What does the Codex skill do and what are its sandbox modes?

It uses the OpenAI Codex CLI for code analysis, refactoring and automated editing. The README lists three sandbox modes: `read-only` for analysis only, `workspace-write` for local file edits, and `danger-full-access` for full permissions including network access.

### Where can I see the changes Code Changelog records?

The skill generates Markdown documentation and serves an HTML viewer in dark mode from a Python server at http://localhost:4000, so changes can be reviewed in the browser while the session is still running.

### What licence is my-skills released under?

The licence field on the repository is not populated, and the project has no GitHub releases. The README does not state reuse terms, so you need to ask the author before copying skill code into a project of your own.

## Sources

- [bear2u/my-skills on GitHub](https://github.com/bear2u/my-skills)
- [Issues](https://github.com/bear2u/my-skills/issues)
- [README](https://github.com/bear2u/my-skills/blob/master/README.md)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/bear2u-my-skills