Find your dream job with Claude Code agents. No coffee breaks, no doomscrolling, no sleep till you're hired!

This project capitalizes on one simple fact: a career search is a numbers game. More real contacts lead to more listings you hear about early, more of those become applications with a person attached, and more of those become interviews. A single human running that chain alone has limited hours in the week. Agents do not. JobFinderOS is a Claude code project that puts agents to work on your search, profiling markets, following companies, crawling job postings, identifying new strategic contacts and preparing you for every interview. Each stage of the job search funnel gets more nurturing than you could feed it yourself, and the judgment stays with you.

This manual covers four things:

1. What it is
2. Getting up and running
3. Using it
4. The methodology: how agentic job search works here

Plus a short reference and a note on contributing at the end.

JobFinderOS is a set of agent personas and skills that run inside Claude Code and work for you to identify career opportunities based on your criteria. They watch the market, find roles, keep your pipeline up to date and honest, and get you ready for every interview. They write everything to a folder of Markdown notes, which you can open in Obsidian or any text editor.

Three agents, each with one job.

       
_{Coach, the recruiter  ·  Scout, the crawler  ·  Mark, the market analyst}

Skills are the commands you run. Each is a short Markdown prompt in .claude/commands/. You type /jobs-daily or /mock-interview in Claude Code and the right agent picks it up. There are 25. Section 3 lists them by when you would use them.

Your config is what they read first. config/profile.md says who you are and what you want. config/scoring_rubric.md says how to score a role. config/recruiter_playbook.md says how the agents behave. Section 4 explains that playbook in plain English.

The vault is where everything lands. vault/Dashboard.md is the daily read. vault/Strategy.md is the weekly one. Every company gets a folder under vault/Companies/ with a profile and one note per role. Digests, market briefs, outreach drafts, and interview prep all have their own folders.

What it never does. It never sends an email or a message. Drafts go to your clipboard and the vault, and you send them from your own client. It never mass-applies. It never sits in the interview.

What it needs. A Claude Code subscription. No API keys. Python is only used by the optional scheduler.

Four steps. The first two happen in your terminal, the last two inside Claude Code.

You need Claude Code installed and logged in with a subscription. No API keys.

claude auth login

macOS / Linux

git clone https://github.com/matthewprice/JobFinderOS.git
cd JobFinderOS
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
claude

Windows PowerShell

git clone https://github.com/matthewprice/JobFinderOS.git
cd JobFinderOS
py -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
claude

The Python setup needs Python 3.11 or newer. Python is only required for the optional scheduler and helper scripts, so if you never plan to schedule anything you can skip the virtual environment and dependency installation and go straight to claude.

Inside Claude Code:

/onboard

This is a 15-minute conversation. It asks about your current role, what you want next, your salary floor, where you will and will not work, companies to avoid, the exact job titles recruiters use for your target, and two or three real wins with numbers. Any career works. It does not assume you are technical.

When it finishes you have four private files, all gitignored:

/jobs-scout

Scout reads your target companies' careers pages, scores what it finds, and writes an opportunity note for anything that clears your bar. Open vault/ and read Dashboard.md. You are running.

Each of these is independent. Add them when you want them.

• Obsidian. Open vault/ as a vault in Obsidian to read the notes with working links. Any text editor works too.

• Gmail. Connect Gmail as an MCP connector in claude.ai. Coach can then triage recruiter email, spot interview invitations, and detect rejections. It is instructed to read only; Section 5 explains what that rests on.

• A schedule. The master scheduler supports macOS launchd and Windows Task Scheduler. Both run scripts/scheduler_tick.py on a repeating interval, every 30 minutes by default. The scheduler reads config/scheduler.yaml and decides whether /jobs-daily, /mark-weekly, or the weekday priority watch is due.

  macOS

  bash scripts/JobFinderOS_install_launchd.sh

  Windows PowerShell

  .\scripts\JobFinderOS_install_windows_task.ps1

  On Windows, the task runs under the logged-in user's session. The computer must be awake for scheduled work to run. Re-running the installer refreshes the task definition.

  A missed run catches up on the next scheduler tick. Runs are logged to logs/ and mirrored to vault/Automation/.

  Check the scheduler without running a Claude Code skill:

  macOS / Linux

  python3 scripts/scheduler_tick.py --dry-run

  Windows PowerShell

  python .\scripts\scheduler_tick.py --dry-run
  Get-ScheduledTask -TaskName "JobFinderOS Scheduler"

  On macOS, you can also verify the local automation setup:

  bash scripts/verify_local_automation.sh

Every skill run ends with a Recruiter's read: two to five sentences of candid advice about what today's state means and what you are avoiding. It is never a recap. Read it.

Coach runs these. /jobs-daily also sends Scout and Mark out first, then hands what they found to Coach.

/jobs-daily     Market pulse and scout run in parallel, then email triage, then the digest.
/checkin        Tick boxes on the Dashboard, tell it what happened, it ages every thread.
/whats-next     "What should we do next?" It sweeps the pipeline and offers the 3 to 5 best moves.

If you run one thing a day, run /whats-next. It is a conversation: pick a move, do it together, the vault updates, the list re-ranks.

Coach runs these.

/warm-path <Company>       Who do you know, or can reach, there? Returns a verdict and dates.
/network-outreach <Co>     Finds 2 or 3 peer-level people and drafts a research-backed note.
/draft-message             A reply, nudge, thank-you, or cold note. Copies to clipboard.
/voice-check               Scans any draft for AI tells. PASS, REWRITE, or DO NOT SEND.

The rule behind this is in Section 4.2. Short version: find a human first, apply two or three days later.

Mark runs these.

/jobs-research <Company>   Business, funding, leadership, product, how they sell, open roles, an angle.
/mark-profiler <Company>   A deeper evaluation for a real decision. Scored, with a verdict.
/title-audit               Reads live postings and tells you what your job is called right now.

Coach runs these.

/jobs-prep <Company> <Role> <Stage>    Full prep brief: company, round, and every person in the room.
/mock-interview <Company> <Round>      Plays the interviewer in character. Scores each answer.
/day-of-card <Company> <Round>         One page to have open: beats, lead stories, questions, traps.
/jobs-cover <Company> <Role>           A cover letter built from research, in your voice.
/story add | find | update             Your career story library. Prep and covers draw from it.

Coach runs these.

/postmortem <Company>      Classifies the objection, logs it in Strategy.md, names the fix.
/profile                   Deepens your wins and voice files over time.

/mark-pulse, /jobs-scout, /jobs-email, and /jobs-digest are the pieces /jobs-daily is made of. Run them alone when you want one piece. /mark-weekly is the weekly brief and Strategy pass, /email-watch is a read-only inbox briefing, and /jobs-priority-watch is the narrow weekday scan the scheduler runs.

This is the operating doctrine the agents follow, written for a person. The source is config/recruiter_playbook.md. Everything here came out of running one real search for five months and studying what worked.

A job board shows you listings. A recruiter tells you what your pipeline means, which thread is dying, which play you are avoiding, and when a pattern has formed. That is the coach's job. The Recruiter's read at the end of every run is the mechanism. If it is uncomfortable, it is working.

An application with no human attached is the last resort, not the default. Before anything is queued, the agents walk a ladder: people you already know at the company, former colleagues who moved there, the hiring manager and their boss by name, a named recruiter. Outreach goes first. The application follows 48 to 72 hours later, so a mention inside the company lands before your resume hits the pile.

When the ladder comes up empty, the application is logged as cold. The system tracks the ratio and aims for 70 percent warm. In the search this was built on, every single rejection was a cold application. Not most. All.

Everyone on the hiring side reads AI-written messages all day and is pattern-matching for them. One detected template can quietly end a conversation, and the people in your field talk to each other. So:

• Low volume by design. At most 3 new people a day, 10 a week. At most 2 follow-ups per thread, then park it for 30 days. Never two similar messages to two people at one company.
• The two-fact rule. Every outbound message carries one fact that took real work to find, with the source cited so you can read it first, and one thing only you could say. Missing either, it does not go. Silence beats generic.
• You send everything. Drafts land on your clipboard with a "before you send" checklist and slots you have to fill in your own words. The agents never send, never schedule, never create a Gmail draft.
• The voice gate. Nothing drafted for you may read as AI-written. No em dashes, no "I hope this finds you well," no perfectly balanced three-part sentences, no flattery openers. /voice-check enforces it.

Every thread is aged against a table of norms. An application at a small company with no reply after 14 days is presumed dead. After a recruiter screen, 7 days of silence means you are the backup candidate, so open a second thread at a comparable company now. A warm contact who has not answered in a week is busy, not hostile, so one bump with a new angle, then park. Presumed-dead threads stop getting your energy. If one revives, that is a bonus, not a plan.

After any rejection, withdrawal, or 30-day ghost, /postmortem classifies the objection, both what they said and what they probably meant: domain-proof gap, level mismatch, location, comp, slate, culture, unknown. It goes into the objection log in Strategy.md. Three of the same objection is a positioning problem, not bad luck, and the fix becomes a strategy action. Nine postmortems in the original search surfaced the pattern that changed its whole back half.

Job titles move, and postings are the trailing indicator. Mark watches for renames at your target companies. /title-audit reads their live postings, tallies the vocabulary, and hands you the exact terms to put in your profile and saved searches. In the original search the target role had been renamed at the companies that mattered. The winning listing arrived three weeks after the search terms changed, under the new name. The crawler never found it. The rename did.

Mass-applying stopped working for everyone at about the same time, because everyone can do it now. The edge moved to the rooms. AI makes serious preparation cheap: a research brief per round, a profile on every interviewer, mock interviews, cover letters that start from research, a story library you know cold. That is where the time saved by the agents should go.

Stages are Applied, Screen, Hiring-manager round, Final, Offer, tracked separately for warm and cold. The weekly Strategy pass does the math out loud. If cold applications are not converting to screens, more cold applications are not the answer. If screens are not converting to hiring-manager rounds, the leak is your narrative. Two more rules from the same section: every serious opportunity gets multi-threaded (recruiter, hiring manager, their boss, a peer) so one silence cannot kill it, and when any process reaches the hiring-manager stage, accelerate the two best comparable ones so offers land in the same ten days. Never bluff an offer that does not exist.

Set a deadline for having an offer in hand. Finals two weeks before that, hiring-manager rounds a month before, screens six weeks before, warm plays now. Every weekly read says in plain terms whether the current pace hits the date and what would change it.

.claude/agents/      coach.md · scout.md · mark.md         the three agents
.claude/commands/    25 skills                             each names its agent in the first line
config/              profile, rubric, wins, voice, targets  your copies are gitignored; templates ship
config/recruiter_playbook.md                               the doctrine in Section 4, in full
scripts/             scheduler, guards, ATS poller, pruner
vault/               the Obsidian vault
CLAUDE.md            project instructions the agents read every run

No code in this repository calls a send or draft API. Gmail is reached only through the claude.ai MCP connector, and that connector does expose send and draft tools. What keeps the agents read-only is doctrine: CLAUDE.md, the agent files, and the playbook all forbid sending and drafting, and every draft goes to the clipboard and the vault instead. Read those instructions before you trust the system with your inbox; there is no separate technical lock.

Your profile, rubric, wins, stories, voice notes, and vault contents are gitignored. Nothing about you ships in this repo, and nothing the agents write goes anywhere you did not put it. If a company you are applying to has guidance on AI use in applications, follow it: you draft first, the coach refines.

Daily digests, run summaries, and daily briefs keep 30 days. Weekly briefs keep 90. On macOS, a weekly launchd helper can prune the rest. The vault is gitignored here, so if you want a permanent archive, back it up to a private repository of your own. Anything worth keeping lives in Strategy.md, Tracking/, or Companies/, never in an old digest.

macOS / Linux

python3 scripts/scheduler_tick.py --dry-run                   # what would run right now
python3 scripts/jobfinderos_run_skill.py jobs-daily jobs-daily # run one skill through the portable runner
bash scripts/JobFinderOS_check_local_runner.sh                # is Claude Code reachable from launchd?
tail -f logs/launchd-runs.log                                 # watch runs

Windows PowerShell

python .\scripts\scheduler_tick.py --dry-run
python .\scripts\jobfinderos_run_skill.py jobs-daily jobs-daily
Get-ScheduledTask -TaskName "JobFinderOS Scheduler"
Get-Content .\logs\launchd-runs.log -Tail 20

The pixel character is Clawd, the Claude Code mascot. He belongs to Anthropic and is redrawn here in work clothes, with affection and no affiliation.

The three agents are the same character in different gear. Coach wears the ball cap and carries the clipboard. Scout has the bucket hat and binoculars. Mark wears the green eyeshade and reads the ticker tape. The drawings live in assets/ as SVG sources; the three agent icons also ship as PNGs because GitHub collapses SVGs inside Markdown tables.

Open any file in .claude/commands/. The first line names the agent. The rest is the task. Copy one, change the task, save it under a new name, and it is a command.

Issues and pull requests are welcome. Bug reports, wording fixes, and "this claim does not match the code" are all useful.

The most valuable contribution is a new skill. Each file in .claude/commands/ is a short Markdown prompt: a description line in the frontmatter, an Agent line saying which of Coach, Scout, or Mark runs it, and a Task section. If you have built one that helped your own search, open a pull request with it. A few things to keep in mind:

• Keep it career-neutral. Skills read config/profile.md for everything about the candidate. Nothing about a specific person, industry, or company belongs in a skill.
• Follow the playbook. Warm path first, low outreach volume, a human in the loop on every message, and a Recruiter's read at the end. config/recruiter_playbook.md is the doctrine; a skill that fights it will not be merged.
• Never send. Drafts go to the clipboard and the vault. No skill may send email, create Gmail drafts, or post anywhere on the candidate's behalf.
• Scrub before you push. Check the diff for your own profile, wins, contacts, or vault notes. The .gitignore covers the usual paths, but a copied example can slip through.

Say in the pull request what the skill is for, which agent runs it, and what it wrote to the vault when you ran it. MIT licensed, so contributions are too.