CareerForge: six agent skills covering job search, resumes, interviews and offer negotiation

Six skills, and how they chain

CareerForge is a pack of six skills for Claude Code or Codex. Each works alone, and they are designed to run in sequence: job-hunt finds roles, resume-match scores a resume against a job description, resume-craft builds or rewrites it, cover-letter writes the letter and the greeting message you send on a recruiting app, mock-interview runs three rounds with per-question feedback, and offer-decision compares offers.

Triggers are slash commands or plain phrases: /job-hunt, /resume-match, /resume-craft, /cover-letter, /mock-interview and /offer-decision.

The sixth is the differentiator the README leans on, and it is the one most job tools leave out. offer-decision compares hourly rate rather than headline salary, works out after-tax take-home, and prices hidden costs, treating a 大小周 schedule as roughly a month of lost leave. It then draws a six-dimension radar chart across economic value, growth value, platform value, track value, quality of life and safety margin, and produces a negotiation package with an anchor price, a floor price and three scripts.

job-hunt: the one that needs the network

job-hunt is the only skill in the pack that needs live web tools. The README marks it as strongly dependent on web_search and web_fetch and states that its core function is unusable without them.

What it does with them is search 30 plus job platforms across ten regions. Seven regions were added on 2026-06-09: Australia and New Zealand, the United States and Canada, the United Kingdom and Europe, Japan, Korea, and Singapore with Southeast Asia. Named platforms include Seek AU/NZ, Reed, Saramin, JobStreet, ZipRecruiter and Daijob.

Results land in an Excel file with a visa column that records whether a posting provides sponsorship, leaves it unmarked, or restricts applications to local candidates. Overseas searches run mainly on English keywords, with Chinese coverage for community job posts.

resume-craft and the two output paths

resume-craft sits at the other end of the pipeline and works entirely on local files. It offers six HTML templates and seven Word templates and can emit HTML, PDF and Word.

The 2026-06-16 update is where it became more than a template picker. Generated HTML resumes gained a browser edit mode, so you click an edit button, change the text in place, and export a PDF without going back to the model. Seven Word templates were added, three Chinese and four English, filled automatically with python-docx to produce editable .docx files. The HTML set was trimmed to six by dropping Clean Teal, and Word templates moved into their own templates/word/ directory rather than sitting alongside the skill code.

Installing it with the Vercel skills CLI

One command installs, and rerunning it updates, because it pulls the latest version from GitHub over the local files:

npx skills add rebecha1227-a11y/CareerForge -g

The installer is built on the Vercel Skills CLI, which the README says detects 50 plus agents including Claude Code, Codex, Cursor, Gemini CLI and Windsurf. Node.js 18 or newer is required. The -g flag installs globally, landing in ~/.claude/skills/ on macOS or the matching .claude/skills/ directory on Windows. Without it, skills go into a .claude/skills/ folder under the current directory and stop working when you move elsewhere. Codex uses ~/.codex/skills/ and Gemini uses ~/.gemini/skills/.

To target specific agents:

npx skills add rebecha1227-a11y/CareerForge -g -a claude-code
npx skills add rebecha1227-a11y/CareerForge -g -a codex -a cursor

A shell installer covers Claude Code, Codex, Cursor, Gemini CLI, Trae, OpenCode and Rovo Dev:

curl -sL https://raw.githubusercontent.com/rebecha1227-a11y/CareerForge/main/install.sh | bash

And there is a manual path:

git clone https://github.com/rebecha1227-a11y/CareerForge.git
cd CareerForge
cp -r skills/* ~/.claude/skills/

Optional dependencies, and what breaks without them

Four Python packages are listed as optional, each tied to one output:

pip install playwright && playwright install chromium
pip install Pillow
pip install python-docx
pip install openpyxl

Playwright generates PDFs in the background, though the README notes you can export from the browser instead. Pillow crops and compresses the resume photo. python-docx fills the Word templates. openpyxl writes job search results to Excel.

Skip one and the matching output disappears rather than the whole skill failing, which is the right way to structure optional dependencies.

Agent and model compatibility, and the DeepSeek caveat

The README is blunt that job-hunt is the only skill with a compatibility problem. resume-match, resume-craft, cover-letter and mock-interview work on anything, because they only handle what you give them or generate local files.

The fully compatible combinations listed are Claude Code with Claude models, which natively support web_search and web_fetch, and Codex CLI with GPT-5.x, where Codex has built-in web search through the OpenAI search cache. Under partial compatibility, Claude Code with DeepSeek V4 is flagged as likely to fail: the DeepSeek API does not support the tool_choice parameter, so web_search and web_fetch calls return a 400 error, with a link to a known issue. The other four skills keep working under that setup.

What the last four updates changed

Four dated entries sit at the top of the README, and they are the best evidence of how the project is being worked on.

On 2026-06-11 job-hunt moved to login-state-first searching. Selecting Boss直聘 or 拉勾 now leads straight into a logged-in session, replacing search-engine snapshots with live data, which the README says increased job volume more than tenfold. Cross-agent browser plugins are supported too, Claude in Chrome for Claude Code and Codex for Codex, and if neither is installed the skill recommends one instead of silently falling back to cookies.

The same update added a three-layer freshness filter: an after: timestamp on the search, a freshness-unconfirmed marker at rating time, and parallel link checks at the quality pass. The README reports that this removed every expired 猎聘 result from a test where the measured staleness rate was 63 percent. Quality checking got four times faster by firing three to five link checks in parallel and no longer verifying amber-listed jobs until you ask. Boss直聘 company and salary scraping selectors were updated, and four skills had bulk content split into references/ for on-demand loading, cutting SKILL.md size by about half on average.

Licence, maintenance and who should skip it

The README carries an MIT badge while the repository metadata reports no licence identifier, so confirm before redistributing. The last push was on 2026-06-16.

The honest limitation is where job-hunt gets its data. It reads search results and, in logged-in mode, pages on commercial job platforms, so it depends on selectors and page structures that change without notice, and on terms of service you should read before automating against them. That the project ships selector fixes as they come up tells you how often they break.

Skip it if you are not on Claude Code or Codex with working web tools. Four of the six skills will still run, but the one that finds jobs will not.

Editorial conclusion

Use CareerForge if you run Claude Code or Codex and want the whole job loop in one place, from searching roles to comparing offers, and you care about the stages past the application: mock interviews, offer comparison and negotiation scripts. Skip it if your agent has no working web tools, because job-hunt is unusable without web_search and web_fetch and the other five cannot find you a job. Install with npx skills add rebecha1227-a11y/CareerForge -g, then type a slash in Claude Code and confirm /job-hunt appears before you go any further.

Frequently asked questions

Official sources

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