jobbot

Walkthrough: The Night Shift, a scroll-driven story of one night of jobbot (18:00 to 09:00) on fictional data, ending in a live demo of its panel (an earlier build, in Spanish). Source: jobbot-showcase.

A local job-search assistant written in Rust. It finds postings from public sources, drops the ones you must not apply to, scores the rest, tailors your CV, fills application forms in a real browser, and keeps a tracker. It is built around one rule: it never says anything you have not told it to say, and it never sends anything without a way for you to stop it.

I built it for my own job search by directing Claude Code: I chose the goals, the architecture and the rules, and reviewed the changes; Claude wrote most of the code and tests (see AI-USE.md). It is a personal tool, not a product: it targets Windows, and part of its output is in Spanish, my native language. The profile and CV in this repository describe a fictional person so that the test suite can run; replace them with your own.

What it does

discover  ->  prepare  ->  submit  ->  digest
 sources      tailored     dry run,      daily summary,
 + filters    CV + PDF     assisted or   tracker, dossiers
 + scoring                 live

1. discover reads public sources (Greenhouse, Lever and Ashby boards, plus RemoteOK, Remotive, Get on Board, Himalayas, Jobicy, We Work Remotely, Working Nomads, 4 Day Week, Workable search, Hacker News "Who is hiring?", python.org and a few others), removes duplicates and excluded companies, checks that the job can be done from your region, scores each posting against profile/criteria.toml, and queues the good ones in SQLite. Postings that expire or turn out to be out of region are withdrawn on the next run.
2. prepare adapts the master CV to each posting: it reorders skills you already listed and adds one sentence. It adds nothing you did not write, and it uses no language model for this. The PDF is rendered with Chromium.
3. submit fills the application form with Playwright. The default is a dry run: it fills what the answer bank knows, saves a screenshot, and sends nothing. --assist fills the form in a visible browser and leaves the click (and any CAPTCHA) to you. --live submits, but only what you approved, and only if the page confirms it.
4. digest writes the daily summary. run-daily chains everything and is what the scheduled task runs.
5. dossier and freelance use Claude to write an interview-preparation dossier or a freelance proposal draft, from the posting and your CV (never from anything else). prep is a spaced-repetition drill for interview questions.

Design rules

These are the parts I would want a reviewer to check.

• Unknown means blank, never invented. The answer bank (profile/answers.toml) answers what you decided to answer. An open question ("why this company?"), a legal question, or self-identification such as gender or ethnicity is left empty and marked in red in the screenshot. --live refuses to click submit while a required answer is missing.
• One work-rights rule lives in code. A posting that is open only because extra_open_places in profile/criteria.toml names its place (Australia in the example) is answered as a move there: "authorized to work there?", "citizen or permanent resident?" and "hold a valid visa?" get "No", "will you need sponsorship or a visa?" gets "Yes", and a visa-status question gets one fixed sentence (bank_for_job in src/pipeline.rs, the answers in src/answers.rs). An empty list turns it off.
• A human reviews the first applications. The first five go through dry run or --assist. Only after that does the program submit on its own, within daily caps and a 90-day cool-down per company.
• Exclusions are tested data. profile/exclusions.toml lists companies, domains and categories the program must never apply to or write about (for example a former employer). They are matched on whole words and domain suffixes, they are stripped from the CV before anything reaches a model, and tests keep it so.
• Idle hours, a kill switch and a lock. Inside the busy hours of profile/schedule.toml the hourly task does nothing, and discover, prepare, submit, run-daily, dossier, freelance and claude-check refuse to start unless given --ignore-window. jobbot stop creates a STOP file: while it exists those seven commands refuse to start (run-daily also looks again between steps, submit before each application), and no flag overrides it. Commands that only keep records (approve, mark, record, task, inbox import, the panel) keep working, and nothing approved under STOP is sent until you resume. A lock prevents two runs from overlapping.
• Postings and e-mails are data, not instructions. They are cleaned and length-limited, and nothing in them is ever executed or obeyed.
• A spending limit on the model. profile/claude.toml caps Claude calls per day and per call. The calls use your own subscription through claude -p, in an empty directory, with no tools, and the program refuses to run if the logged-in account is not the one you expect.
• No CAPTCHA solving, no bot against sites that forbid it. LinkedIn, Upwork and similar platforms are never automated; Ashby forms (which flag automated browsers) are filled for you to review and sent by hand.

Architecture

A Cargo workspace in layers, with the dependency rule enforced by a test (tests/architecture.rs) rather than by convention:

crate	layer	holds
crates/jobbot-domain	0	the pure core: job model, application state machine, scoring, exclusion rules, outbound-text filters, parsing of profile/*.toml, the run window. No I/O, no async runtime, no database.
jobbot (root)	1	the application: SQLite tracker, sources, the Claude client, forms (Playwright), the local panel, the CLI.

The domain decides (is this excluded, what does it score, which state may follow which); the application acts. A mistake in the first cannot send an application. Details in docs/engineering/architecture.md.

Quick start

Requirements: Rust 1.95 (pinned in rust-toolchain.toml), and on Windows the Visual Studio Build Tools with C++.

cargo build --release
target\release\jobbot.exe install-browsers   # downloads Chromium once, about 250 MB
cargo test --workspace                       # the suite runs offline against fixtures

Then make it yours:

1. Replace cv/cv-master.md with your CV (keep the headings).
2. Fill profile/answers.toml. Every value in it is something the program will say on your behalf.
3. Edit profile/criteria.toml (pay floor, threshold, titles, skills), profile/exclusions.toml, profile/sources.toml and profile/schedule.toml.
4. Optionally run jobbot claude-login for the model-backed features; put the expected account in profile/claude.local.toml (git-ignored).

Commands

jobbot discover [--rescore]     find, filter, score and queue postings
jobbot top --n 20               review the best-scored postings
jobbot prepare                  tailor the CV per posting and render PDFs
jobbot submit [--assist|--live] dry run by default; --assist fills a visible browser; --live sends approved ones
jobbot approve KEY...           approve dry-run applications
jobbot resolve KEY...           put a "needs you" application back in the queue
jobbot confirm KEY...           mark as sent something you sent by hand
jobbot digest                   daily summary
jobbot run-daily                discover + prepare + submit + dossiers + digest
jobbot status                   counts by state, the last run, the verified Claude account
jobbot results [--last 20]      latest submit results (mode, state, answers filled, what was missing)
jobbot stop | resume            kill switch
jobbot mark KEY STATE           record a reply by hand (replied, interview, offer, rejected, ...)
jobbot record ...               record an application you sent yourself
jobbot inbox import FILE        import job e-mails summarised elsewhere (the program never touches your mail)
jobbot today                    your list for today
jobbot task add TEXT | done ID  add a to-do to today's list, or mark one done
jobbot dossier KEY...           interview-preparation dossier (calls Claude)
jobbot freelance profile|proposal PLATFORM   draft only; nothing is posted anywhere
jobbot prep ...                 spaced-repetition interview drill
jobbot panel                    local web panel (applications and review)
jobbot claude-check | claude-login
jobbot install-browsers

Inbox file

jobbot inbox import FILE reads a JSON array of at most 500 messages (2 MB). Each message needs message_id, thread_id, from, subject, received_at (RFC 3339), kind (ack, rejection, recruiter, interview_invite, technical_test, action_required or other) and summary; company, job_key and task ({"text": ..., "due_at": ...}) are optional. Every text is cleaned and cut to a fixed length, and a message already imported is skipped.

Scheduling

Scheduling on Windows is two scripts: scripts/schedule.ps1 registers an hourly task, \jobbot\run-daily, that exits immediately outside the run window, and scripts/unschedule.ps1 removes it. Neither runs by itself.

Tests and tooling

• 1,175 tests run offline with a plain cargo test: sources against recorded fixtures, forms against local HTML pages, Claude against a fake claude executable. Another 109 are marked #[ignore]: 107 need a real Chromium (form filling, PDF rendering, the panel; two of them act only when JOBBOT_SCAN_URLS / JOBBOT_REAL_URLS list live pages), one asks a logged-in claude (a real call) and one queries the Windows Task Scheduler, read-only. After install-browsers, cargo test -- --ignored --test-threads=1 --skip real_claude --skip real_scheduler runs the browser ones.
• CI (.github/workflows/ci.yml) runs cargo fmt --check, cargo clippy -D warnings and cargo test on Windows (and, informationally, Linux), plus cargo deny and cargo audit. Actions are pinned by commit.
• unsafe_code is denied for the workspace and dbg!, todo! and unimplemented! are denied by clippy.
• About 60,000 lines of Rust in 81 files, tests included.

Limits, honestly

• Windows first. The scheduler scripts are PowerShell; Linux is a CI job that is allowed to fail.
• Much of the CLI and panel text is in Spanish, and so are the interview-practice questions in src/prep/questions.toml and the CLI help (the doc comments in src/cli.rs). The example profile and the scripts are in English, as are most code comments.
• Times are Santiago (Chile) time: the CLI, the digest, the dossiers, the freelance drafts and the panel show them in that zone, and the dates of jobbot today and of the digest follow it. The zone is fixed in the code; only the busy hours in profile/schedule.toml take any IANA zone.
• The first build downloads the Playwright driver (Node.js from nodejs.org and playwright-core from npm) into the user cache. playwright-rs checks the TLS connection but no checksum; to build against a driver you checked yourself, point PLAYWRIGHT_DRIVER_CACHE_DIR at it.
• Greenhouse, Lever and Ashby forms are handled; other application systems are listed for you to open by hand.
• The local panel's page is a single built HTML file (src/panel/page.html); its front-end sources are not part of this repository.
• This is the code of a tool one person uses on their own accounts. It is published so the engineering can be read, not as a service.

License

MIT, except the third-party fonts, code and icons listed in THIRD_PARTY_NOTICES.md, which keep their own licences (the fonts: SIL Open Font License 1.1, texts in licenses/fonts/).