career-agents

Single pipeline for the job search in out/pipeline.md with stages found, matched, applied, screen, interview, offer, closed, follow-up dates and a daily "what is due" list. Written to by job-scan, job-apply, recruiter-reply, interview-prep and offer-review. Use when the user asks what is due, where things stand, or to update a stage.

Orchestrator. Run the full LinkedIn + CV + personal website sync for one person, or resume it at any stage. Use when the user says "sync my profile", "update my LinkedIn and CV", "run the career flow", or asks what is left to do.

One-time setup for a person. Installs the LinkedIn MCP server, checks uv, poppler and Chrome, creates the private career workspace from templates and walks through career.json. Use when someone new wants to use the flow or when verify_kernel fails.

Write a case study for the personal website and the matching LinkedIn Project entry from one role or project in the canonical CV, using the pattern "what the product is, what I did in order, stack, outcome". Use when the user wants a project page, a portfolio entry, or richer Projects on LinkedIn.

Write a one-page cover letter or short application note for a specific job from the canonical CV, positioning and the job description. Use when the user shares a job link, a job post or a company name and wants to apply. Never sends anything.

Apply accepted audit findings to the CV through the configured adapter (Canva, DOCX, Markdown/HTML, Google Docs), export a PDF and verify it. Use after the user has chosen which findings to apply. Every commit and export needs an explicit yes.

A/B test LinkedIn headline (or About first line) variants across two-week windows using search appearances and profile views from weekly-review, then keep the winner. Use in assisted or auto mode when the user wants evidence rather than opinion on the headline.

Read recruiter email and calendar invites through the user's Gmail or Outlook and Google Calendar MCP servers, turn them into pipeline rows, trigger interview-prep for invites with the right date, and put reminders in the calendar. Extends the LinkedIn inbox reading to the channels recruiters actually use. Use daily in cron or when the user says "check my email for recruiters".

Guided onboarding interview that fills a new person's whole workspace kernel from a conversation, their CV PDF, their LinkedIn and three things they wrote. Infers voice from real writing. Use for a new person after career-setup, or when verify_kernel reports placeholders.

Right after an interview the user dictates or pastes their notes; the skill writes the thank-you note, logs the real questions asked, scores how it went, updates the mock-interview bank and the prep file, and moves the pipeline row with the next step and due date. Use within 24 hours of any interview.

Prepare for a specific interview stage. The user pastes the invite email or LinkedIn message, or the skill reads it from the LinkedIn inbox, or they just say "Tech interview 2 at Acme". It detects the stage (HR screen, hiring manager, tech 1, tech 2, system design, EM, PM or PO, CTO or Head of Engineering, final, panel), builds a stage-specific brief with STAR answers from real CV facts, researches the interviewers and company, and hands over to mock-interview. Use when an interview is scheduled or the user asks to prepare.

Search LinkedIn jobs through the MCP server, open every candidate's full description and run the fit gates (role, location and work authorization, hours, language, real job, applicant volume) before recommending anything. Use when the user says "scan jobs", "find jobs", "today's jobs", "what's new for me". Never applies or messages.

Turn the skills-gap output into a six-week, project-based plan whose deliverable is real proof (a repo, a case study, a post series), with weekly checkpoints and the hours the person actually has. Use after skills-gap, after a tech-stage rejection, or when the user asks how to learn X for the target role.

Design and render a LinkedIn cover banner (1584x396) as PNG from an HTML template using headless Chrome. Use when the user wants a banner or the audit finds the current one is missing or off-message.

Draft LinkedIn posts in the person's voice from real work and CV facts, track what was already posted, and suggest topics that do not repeat the last series. Use when the user says "draft a post", "what should I post", "LinkedIn post", or a calendar slot from career-activation is due. Never publishes.

Produce paste-ready LinkedIn copy (headline, About, experience descriptions, projects, skills, featured, certifications) from the canonical CV and positioning, in rounds with a "what is left" list. Use after an audit or when the user wants a stronger LinkedIn profile.

Interactive interview practice in chat for a named stage (HR screen, hiring manager, tech 1, tech 2 or system design, EM, PM or PO, CTO or Head of Engineering, final panel). The user pastes the invite or says "mock interview, EM round at Acme"; the skill plays that interviewer, asks one question at a time, scores each answer against STAR and the CV facts with one fix, then summarises. Use after interview-prep or whenever the user says "practise" or "mock interview".

Produce a second-language version of the headline, About, experience descriptions and CV (German, French, Spanish, Portuguese, or any language the user names) that stays fact-identical to the canonical CV, with a back-translation check. Use for people targeting a local market or when a job post is in another language.

Mirror the workspace outputs (pipeline, briefs, debriefs, weekly reviews, growth plan) into the person's own notes tool, Notion or Obsidian, through its MCP server or the vault folder, one way from the workspace, so they read and annotate where they already work. Use when the user keeps notes in Notion or Obsidian.

Send the outcome of a run as a short digest to the channel the person configured (Slack, Telegram, email, or a file) through the matching MCP server, so they read a summary instead of opening the workspace. Used by career-cron, weekly-review, job-scan and application-tracker; also on demand.

Three-way audit of LinkedIn vs CV vs personal website. Runs the word-level differ, then the read-only profile-differ agent classifies every difference as FACT, WORDING, ACCEPTED or NOISE and lists CV quality issues. Use before any rebuild or CV update, or when asked "do my profiles match".

Snapshot every LinkedIn field before the operator edits it and after, keep the versions in the workspace with a diff, and restore a field to any previous version through linkedin-apply. Safety net for automated edits and a record of what changed when metrics moved. Runs automatically inside linkedin-apply; also on demand.

Log every rejection with stage, reason and source, then find the pattern over the last 30 and 90 days (always out at the tech screen, never past HR on range, no replies to cold applications) and name the fix and the skill that applies it. Runs weekly in cron and on demand.

Which target countries and remote setups are realistic given the person's work authorization, what job posts in each market say about sponsorship and remote-in-country rules, and how to phrase location and authorization on LinkedIn, the CV and applications. Information from job posts and official pages the user provides; not immigration advice.

Build a salary range table for the target role and markets from job posts that state a range (LinkedIn search results and job details through the MCP, plus posts the user pastes), with sample size and currency, feeding positioning.md and offer-review. No paid data, no guesses; small samples are labelled small.

Audit and update a personal website so its facts, CV download, chatbot knowledge base and metadata match the canonical CV. Uses the read-only site-auditor agent first, then proposes a patch in the site repo. Commit and push only with an explicit yes.

Compare the skills demanded by recent target job posts with the skills on the CV and LinkedIn; report what to surface (already known, not visible), what to learn, and what to drop. Use after a few job scans, or when the user asks what to learn next or which skills to add to LinkedIn.

Conference and meetup call-for-papers matcher with abstract drafts, plus an article calendar for dev.to, Medium or the personal site that reuses content-engine output. Use when the user wants speaking or long-form writing as a career signal.

Script and shot list for a 60-second intro video or LinkedIn cover story, from the pitch in positioning and the strongest CV facts, in the person's spoken voice, with captions text and a thumbnail card. Use when the user wants a video on LinkedIn, the site, or for an application that asks for one.

Weekly feedback loop. Reads LinkedIn analytics (search appearances, profile views, followers, post impressions, inbound messages) through the browser operator, or from pasted numbers in draft mode, logs them in out/metrics.md, compares with previous weeks and says what to change next week. Use on the review day or when the user says "weekly review" or pastes analytics numbers.