AI Agent Skills
Flowxtra ships two complementary skill sets so AI assistants can run hiring workflows for you in plain language. They look similar but serve different runtimes — pick the one that matches how your team works.
Quick comparison
| CLI-driving skills | Claude-native skills | |
|---|---|---|
| Repo | Flowxtra/flowxtra_cli (in skills/) | Flowxtra/flowxtra-claude-skills |
| Runs in | Claude Code, Cursor, Codex, any agent supporting npx skills add | Claude.ai web, Claude Desktop |
| Drives | The flowxtra CLI binary | The Flowxtra MCP server tools |
| Prerequisite | CLI installed (npm install -g @flowxtra/cli) + flowxtra auth login | MCP connector added in Claude settings |
| Install | npx skills add Flowxtra/flowxtra_cli | Upload .skill files in claude.ai → Settings → Capabilities |
| Best for | Developers in a terminal/IDE | Recruiters using claude.ai in the browser |
Both sets cover the same core workflows. You can install either, or both — they don't conflict.
CLI-driving skills
For agents that can shell out to the flowxtra command. Three focused skills, plus the
core flowxtra skill that teaches the agent the full CLI surface.
| Skill | Triggers on |
|---|---|
flowxtra | Any request involving Flowxtra. Teaches the agent the full CLI. |
post-job | "create a job", "publish a vacancy", "post a role on Flowxtra" |
screen-candidates | "review applicants", "shortlist top N", "move [name] to interview" |
Install:
# 1. The CLI (skills drive it)
npm install -g @flowxtra/cli
flowxtra auth login
# 2. The skills
npx skills add Flowxtra/flowxtra_cli
Then say "post a senior backend job" and the agent runs flowxtra jobs create … for
you, asking for what it needs along the way.
Claude-native skills (MCP-based)
For users of claude.ai or Claude Desktop — no CLI install, no terminal. Five specialized skills that drive the Flowxtra MCP tools directly.
| Skill | Triggers on |
|---|---|
flowxtra-post-job | "create a job", "post a vacancy on Flowxtra" |
flowxtra-screen-candidates | "show applicants for…", "shortlist", "move to interview", "reject" |
flowxtra-promote-job | "announce on LinkedIn", "share the role on social" |
flowxtra-schedule-interview | "book an interview", "schedule a call with the candidate" |
flowxtra-hiring-report | "how many applicants", "channel breakdown", "pipeline status" |
Install:
- Connect the Flowxtra MCP server in Claude — see MCP Setup.
- Download the
.skillfiles from the releases page or zip each folder yourself. - In claude.ai → Settings → Capabilities → Skills, click Upload and select each
.skillfile.
That's it — the next time you ask claude.ai "post a senior dev job on Flowxtra" the right skill fires automatically.
What each one does
- Post a job: drafts the listing first, shows you the copy, publishes only on confirmation. Optional follow-up to add custom application questions and announce on social.
- Screen candidates: lists applicants, surfaces who looks strongest, moves them through your pipeline. Always confirms before sending stage emails or rejections — they go to real people.
- Promote a job: writes a tailored caption for each connected social channel and
posts after your OK. (Note: branded image generation is a CLI feature —
flowxtra jobs promotegenerates 13 variants; the MCP version posts text + apply link.) - Schedule interview: picks the candidate, confirms time/timezone, books via Google Meet / Zoom / Jitsi.
- Hiring report: read-only — answers questions like "how many applicants this week", "which channel performs best", "where are candidates stuck in the pipeline".
Safety
Every outward-facing action (publishing a job, posting to social, sending an email, moving or rejecting a candidate, booking a meeting) is confirmed with you before it runs. The skills surface the change first, then act on your OK.
Which should I use?
- You live in a terminal or IDE (Claude Code, Cursor, Codex) → CLI-driving skills. The CLI is faster, scriptable, and includes features the MCP can't expose (like the 13-variant social-image generator).
- You use claude.ai in the browser → Claude-native skills. No install. Pair with the MCP server.
- Both is also fine — they target different runtimes.