CareerForge, six skills and one that needs the internet

Five skills work offline and one cannot work at all

The compatibility section is the most useful thing in the repository, and it is a table of what each skill needs.

Four of them need nothing. resume-match processes only the resume and job description you supply. resume-craft generates local files. cover-letter is text generation. mock-interview is a conversation. Each is marked as fully usable in any model and agent combination.

One is different. job-hunt is marked as strongly dependent on two specific tools, web_search and web_fetch, with the note that its core function is unavailable without them. There is no degraded mode described and no offline subset, which is the honest framing: a job search that cannot reach the network is not a job search.

The practical reading is that installing this toolkit is cheap even if the model situation is awkward, because four fifths of it will run regardless. The cost of a bad model choice lands entirely on the one skill you cannot use offline.

The table lists five skills and the toolkit has six, so the newest one has no compatibility entry. That is a small gap and an easy one to fill, and it happens to be the skill that reads job offers rather than the ones that read job ads.

The failure modes are named, not hand-waved

Where the toolkit does not work, the documentation says why, and the reasons are specific.

Claude Code with a third-party model is listed as likely to fail, with two separate causes named. DeepSeek is called out individually: the API does not support the tool_choice parameter, so web_search and web_fetch return a 400, and the README links an upstream issue. Other third-party models fail for a broader reason, which is that most Anthropic-compatible endpoints do not fully implement the tool-calling protocol.

Codex with a third-party model is listed as unusable rather than degraded. The cause is a protocol mismatch: Codex requires the Responses API and DeepSeek only supports Chat Completions, so a conversion proxy such as OpenRouter or Moon Bridge is needed, and the README says tool-calling stability through that proxy is not guaranteed.

The two recommended combinations are Claude Code with Claude models, called the first choice for native support of both tools, and Codex CLI with the GPT-5.x series, where search goes through the provider's own search cache.

There is also a suggested workaround if you want the cheaper model: run job-hunt on a native model and the offline four on the third-party one. That advice is more useful than a flat warning, because it isolates the constraint instead of telling you to change your setup.

One command, auto-detected across fifty agent types

The installation story is built around a single command and an explicit claim that you do not need per-agent instructions.

npx skills add rebecha1227-a11y/CareerForge -g

That is based on the Vercel Skills CLI, and it auto-detects which agents you have installed, with the README naming Claude Code, Codex, Cursor, Gemini CLI, and Windsurf among more than fifty. It requires Node 18 or newer.

The global flag is the recommended form, with the reason given plainly: it means the skills work whichever folder you open the agent in. Without it, the skills land in a .claude/skills directory under the folder you happened to be in, and the note says you will not be able to use them from anywhere else. Global paths are given per platform, and the equivalents for Codex and Gemini are named too, along with the claim that the CLI places files correctly on its own.

You can also target agents by name with a flag, one or several at a time. Verification is described as typing a slash in the agent and seeing the commands, or just asking for the thing in natural language.

Three other routes exist: a shell installer piped through bash, a manual clone and copy into the skills directory, and handing the repository link to an agent and asking it to install the package.

A 63 percent expiry rate is why the search filters three ways

The June update notes contain the most honest numbers in the repository, and they are about failure.

The claim is that a measured expiry rate as high as 63 percent was found in results from one job board, and that all of those results were removed. The response was a three-layer recency filter rather than a single fix.

The first layer appends an after timestamp to the search itself, so the query is bounded before results arrive. The second marks listings whose recency could not be confirmed during rating, rather than silently treating them as fresh. The third verifies links in parallel during quality checking, and only for listings that passed the earlier stages.

That ordering is the interesting part. Verification is the expensive step, so it runs last and only on survivors, which is also why the notes say quality checking got four times faster: link verification now goes out three to five at a time, and mid-confidence listings are no longer checked proactively at all, but only when you ask about them.

A separate change in the same release is worth noting for how it was handled. When Boss Zhipin or Lagou is selected, the tool now goes to a login-state search instead of a search engine snapshot, which the notes say raised the volume of listings by more than ten times. And when the browser plugin is missing, it does not fall back quietly: it recommends installing the plugin and then asks whether to proceed with cookies instead.

A job board changed its DOM and the scraper broke

One line in the changelog captures the maintenance reality of anything that reads another company's web pages.

Company name and salary extraction from Boss Zhipin stopped working, and the fix was to update the JavaScript selectors to match the current page markup, plus add a full-text fallback parse so the next markup change does not produce another silent failure.

Both halves of that fix are worth noting. The selector update is the part anyone can predict. The fallback parse is the part that matters, because it means the tool degrades to a cruder extraction rather than returning nothing, which is a different failure mode for whatever consumes the output.

This is the same shape as the login-state change described above. Both are recognitions that a search built on scraped search-engine snapshots was both slower and worse than going to the site directly with a session, and the 63 percent expiry figure is the number that probably motivated the change.

The other item in that release is invisible from the outside and useful anyway. Large sections of four of the skills were moved into references directories that load on demand, which cut the average skill definition file in half. A token saving of that size changes how much of the toolkit can sit in context at once.

The Word templates moved out of the skill directory

The resume skill is the largest of the six, and its file layout is documented in detail.

It has a skill definition file, a templates directory holding six HTML templates with previews plus reference templates, a scripts directory for PDF generation, photo processing, and Word filling, and a references directory with design specifications. The other skills are much thinner: job-hunt and resume-match have a definition plus references, and cover-letter has only a definition file.

The June change to that skill added two things. An edit mode for HTML resumes, so you can click into the generated page in your browser, change the text yourself, and export a PDF, with the stated benefit that small edits no longer need the model at all. And seven Word templates, three in Chinese and four in English, filled automatically by a Word document library into editable files.

The same release also removed one HTML template rather than adding one, taking the list from seven to six, with the reasoning given as removing a design that was less distinctive than the others.

One structural change is easy to miss. The Word templates were moved into a top-level templates directory rather than living with the skill code. That matters because the skill directory is what gets copied into your agent's skills folder on install, so anything left there is copied on every update, and templates are among the bulkiest files in the repository.

There is a licence badge and no licence file

The repository root is small enough to audit line by line, and it has no licence file in it.

What is there: a git ignore file, an architecture document with a Chinese filename, the README, a docs directory, the shell installer, the skills directory, a templates directory, and a test-output directory. There is no LICENSE, and the project's licence is not recorded anywhere.

The README does carry a licence badge near the top, pointing at a file that does not exist in the repository. So a reader arriving from the README badge gets a broken link, and the licensing claim in the README text, that the project is open source and free, has no document behind it.

The test-output directory is the other thing worth noting. Output from a test run is checked into the repository rather than ignored, which is normal for a small project and produces a tree that includes artefacts nobody chose to keep.

Release history is empty as well. There are no GitHub releases, and the last push to the main branch is 16 June 2026, which is the date of the newest changelog entry as well. For a project distributed by copying files into an agent's configuration directory, there is no install to pin, so the changelog entries are the version history.

Editorial conclusion

CareerForge is worth installing globally if you use Claude Code or Codex and you are job hunting now, since four of the six skills are offline and the sixth is the only one with a compatibility question attached. Check four things first. Whether your agent has working web search, because job-hunt is documented as unusable without it. Whether you want the third-party model at all, since the recommended path is Claude Code with Claude or Codex with GPT and the alternatives come with named failure modes. Whether you have Node 18 or newer for the recommended installer. And whether you intend to redistribute anything, because the repository carries a licence badge and no licence file.

Frequently asked questions

Official sources

1. Issues
2. README
3. rebecha1227-a11y/CareerForge on GitHub