Imported from CervantesVive/job-search-copilot (plugins/job-search-copilot/skills/job-search-setup/SKILL.md). Install upstream with npx skills add CervantesVive/job-search-copilot --skill job-search-setup. Copyright stays with the author.

Job search setup

Runs once per person, at the start. Conversational, one question at a time — the person using this may not be technical, so keep every message short, plain-language, and free of jargon (no "cron", "artifact", "MCP" in anything you say to them — just "your tracker page", "check every few days", etc.).

Read ${CLAUDE_PLUGIN_ROOT}/shared/conventions.md before doing anything else in this skill — it defines every file name and data shape referenced below.

0. Check for an existing setup

Call project_memory_read with no arguments. If tracker-data.md already exists, this person is already set up: read tracker.md for the link (if a webpage was ever published), tell them their tracker already exists, give them the link again, and ask if they want to update their target profile instead (roles, locations, companies) rather than starting over. If they do, jump to step 2 and only touch profile.md / search-config.md, not the tracker data structure. Otherwise stop here.

1. Resume intake

If a resume file is attached to their message, read it directly. If not, ask for one first, in plain terms: "Can you upload your resume? I'll use it to figure out what roles and companies make sense to search for." Wait for it.

From the resume, extract a short working summary — current title/level, years of experience, core skills, 2-3 standout achievements or companies. You will write a condensed version of this to profile.md in step 3, not the full resume text.

2. Interview

Ask these one at a time, waiting for each answer before moving on. Skip a question if the resume or a prior answer already makes it obvious, and say so ("Looks like you're targeting product management roles based on your resume — sound right?") rather than asking blind.

1. Target role(s) and level — job titles and seniority to search for.
2. Industries or company types — specific industries, or "open to anything."
3. Locations and work style — cities/regions, and remote / hybrid / onsite preference.
4. Specific companies — any dream companies or ones they already have an eye on. These seed the tracker's watchlist. Fine to have none.
5. Deal-breakers — anything that's a hard no (e.g. no relocation, no 5x in-office). Optional.
6. Job boards to search — propose sensible defaults for their role/ location (e.g. "LinkedIn Jobs and Indeed — I'll search those with your target titles and location"). Ask if they have others, or a saved search URL they already use, and want it included.
7. How often to check — offer a default and let them accept it rather than making them design a schedule: "I'll check job boards every couple of days and work through company career pages in rotating batches every few days, so nothing gets missed but I'm not hammering the same sites constantly. Want me to check more or less often than that?"
8. LinkedIn network matching (optional) — explain in plain terms: "If you want, I can flag when a new lead is at a company where you already know someone. That needs a one-time export of your LinkedIn connections — go to LinkedIn → Settings & Privacy → Data privacy → Get a copy of your data, request just 'Connections', and LinkedIn will email you a link in a few minutes. Whenever it arrives, come back here, drag the Connections.csv file into the chat, and say 'here's my LinkedIn export.' Want to set that up now, later, or skip it?" If they have the file already or get it during this conversation, process it per step 5 now; otherwise move on and handle it whenever it shows up in a later message (any skill in this plugin should recognize an uploaded Connections.csv and process it per step 5, then confirm to the user).
9. Gmail reply tracking (optional) — explain: "I can also watch your email for recruiter replies and flag ones you haven't responded to yet. Want to connect Gmail for that?" If yes, use SuggestConnectors with keywords ["gmail", "email"] and let them complete the connection. If no or they skip, note it and move on — this can be turned on later just by asking.

3. Write project memory

Write profile.md, search-config.md (job_boards, companies list seeded from step 2.4 at priority: high plus any others they mentioned, cadence from 2.7, rotation_index: 0), and gmail-status.md (connected: yes/no) per the shapes in shared/conventions.md. Each with proper frontmatter. Update MEMORY.md's index with one line per file.

4. Create the tracker

Write tracker-data.md per the JSON shape in shared/conventions.md — this is the real tracker, and it exists independent of any webpage. Seed watchlist from any companies named in step 2.4 (empty applications and leads — the scans will populate those). Use the person's first name and a one-line profile summary as profileSummary.

Then run the sync procedure from shared/conventions.md to publish the first webpage view (this is an interactive session, so this should succeed normally) and write tracker.md with the resulting URL.

5. Process a LinkedIn export, if provided

If Connections.csv was uploaded (now or, per step 2.8, in some later message), read it. LinkedIn's export has a header row (commonly First Name, Last Name, URL, Email Address, Company, Position, Connected On, though exact columns can vary — read the actual header row rather than assuming). For each row, keep only Company, full name, and Position. Skip rows with no company. Write the condensed list to network.md. Confirm to the person how many connections were imported and that new leads will now get flagged when they're at a company where this shows a contact.

6. Schedule the recurring scans

Use create_trigger (never a local/in-session cron tool — those don't survive between sessions). Create:

• Job board scan — cadence from step 2.7 (default: every 2 days). Prompt: "Run the job-search-scan skill's job-board mode for this project's job search (job-search-copilot plugin). Read shared/conventions.md and search-config.md first, then search the configured job boards, dedupe against the tracker, add new matches to the Leads section, and end with a short plain-language summary of what's new."
• Company career page scan — cadence from step 2.7 (default: every 3 days). Prompt: "Run the job-search-scan skill's company-page mode for this project's job search (job-search-copilot plugin). Read shared/conventions.md and search-config.md first, work through the next batch of companies starting at the stored rotation_index, update rotation_index when done, dedupe against the tracker, add new matches to the Leads section, and end with a short plain-language summary."
• Gmail reply check, only if Gmail was connected in step 2.9 — daily. Prompt: "Run the job-search-gmail-check skill for this project's job search (job-search-copilot plugin) and end with a short plain-language summary of anything needing a reply."

Set notifications: {"push": true} on each so the person gets pinged when a run finds something, and initiation: "human_request" since this was set up at their explicit ask. Leave environment_id unset so each trigger runs in this project's environment.

7. Prime it now

Don't leave the tracker empty until the first scheduled run. Right now, in this conversation, do the same work job-search-scan and job-search-gmail-check would do on their first scheduled firing:

• Run job-board mode and company-page mode (one batch, same rotation logic as always — don't special-case a bigger or smaller batch for this) against search-config.md.
• If Gmail was connected, also run a gmail-check pass.
• If job_boards and companies are both empty (the person skipped naming any), skip this step entirely and say so plainly in the wrap-up rather than pretending to have scanned nothing.

Follow those skills' own logic for writing tracker-data.md and syncing the webpage (per conventions.md) exactly as a scheduled run would — this is not a special mode, it's those skills' normal behavior, just invoked now instead of waiting. Don't have them send their own standalone summary reply, though — hold the results (how many leads found, anything needing a reply) to fold into the wrap-up in step 9, so the person gets one cohesive message instead of several back to back.

8. Optional: keep the tracker webpage always current

Your data is always safe regardless of this setting — every scan writes to tracker-data.md directly, not the webpage. This setting only affects whether the webpage re-renders itself automatically right after a scheduled scan, versus catching up the next time you're chatting here (which happens automatically too, just not instantly). Mention it as a minor, skippable option, not a required step:

"Optional: if you want your tracker webpage to always show the very latest the second a scheduled scan finds it — rather than catching up next time you're chatting here, which happens either way — there's a one-time setting for it. Open the cloud/environment icon near the message box, edit the environment in use (probably 'Default'), and set Network access to 'Full' or 'Custom' with *.frame.claudeusercontent.com allowed. Totally optional — skip it if that sounds fiddly, nothing about your data depends on it."

9. Wrap-up message

Fold in what the priming pass in step 7 actually found — "I already went ahead and ran your first scan: found 3 leads, added them to your tracker" (or "nothing matched yet — I'll keep checking" if it came up empty, or "skipped this since no boards or companies were set" if step 7 had nothing to work with) — rather than only describing what will happen later. Then cover: the tracker link, what's scheduled and roughly how often (scheduled scans catch up whenever you next open Cowork, rather than firing to-the-minute in the background — so checking in every few days gets you the freshest results), and three things they can just say any time — "prep me for an interview at [company]", "practice interview for this posting" (a job link or pasted description), and "I applied to [company]" or "add [company] to my watchlist" to update the tracker by hand. Keep it to a few sentences plus the priming-results line — they now have a working, populated link to check instead of a wall of text.

Also work in, low-key, one more sentence — not its own callout, just another line among the others: "One more thing: from now on, each automatic check runs as its own new chat that'll show up in your sidebar under 'Scheduled' — that's totally normal, you never need to open them, your tracker page is always the up-to-date summary."