Job Search Panel AI-⁠assisted job hunting on hh.ru

Saved hh.ru searches are checked on a schedule. New vacancies go through the hard rules of the questionnaire — stop words, salary, region, work format — and get a preliminary score without any AI, from the weights of the skills.

Above the board is a 30-⁠day funnel: how many vacancies were found, scored, sent and viewed, and how many invitations came. The working list moves with J/K, Enter opens the application on hh.ru, S skips a vacancy and D puts it off.

The panel packs suitable vacancies into batches of 12 in an exchange folder. Claude works through them in a Claude Code session with the /hh-⁠batch command: it reads the profile and the rules, writes the result against a JSON schema, and the panel imports it by itself. The product makes no AI API calls and holds no API keys.

The review can also run without a person: the panel starts claude -p itself when batches appear and reports to Telegram how many were reviewed and how many fit. The background run has narrow rights: it writes only to the result folders, the only command it may run is the style checker, and it can’t read .env, the database or the backups.

To Claude the vacancy texts are data, not instructions: letters with fragments that look like secrets are rejected on import. Claude writes Russian letters by the humanizer-⁠ru skill (MIT, by Ilya Utov), and the panel checks imported and edited letters with the same scanner.

On 15 December 2025 hh.ru restricted its API methods for job seekers, and what a given app may use is not known in advance. So the panel runs a probe and picks a mode from its results: if applying through the API is closed, Assistant mode takes over; if the list of applications is closed, a manual log and hh.ru emails read over IMAP.

Every request to hh.ru goes through one client: OAuth with PKCE, request rate, retries, errors, token refresh and captcha. Methods missing from the public specification are called only after the probe confirms them. The panel keeps to hh.ru’s limits and its own stricter ones — per day, per hour and a sending window — doesn’t emulate a browser or get around captcha, and asks for consent before the first application: applications go out in the user’s name.

In Telegram, events of one kind within five minutes are merged into one message. A vacancy with a letter comes with the buttons “Yes, sent”, “Skip” and “Put off”, and the bot takes commands only from the owner’s chat.

A segment is any slice of the base: a saved search, a criteria set, a target title, a work format, experience or a tag. Tags — industry, tasks and tools — are set by the panel from the title, skills and description of a vacancy, Claude adds its own when scoring, and labels that mean the same thing count as one.

“Hot” vacancies are fresh, well scored and have few applicants, so they are worth opening first. “Borderline” ones fell a few points short of a threshold and are worth a look by eye.

The summary counts conversions from a found vacancy to an invitation, the median employer response time and the daily trend, and any data exports to CSV. The reports are computed in a separate thread, so the rest of the panel doesn’t freeze meanwhile.

Channels are read through the public web version at t.me/s, without signing in to Telegram. The panel reads company and studio sites without a browser, respects robots.txt and doesn’t get around bot protection.

Tenders are checked once an hour and compared with the questionnaire. Studios found through hh.ru are offered only an unsolicited application: the hh.ru agreement allows its data to be used for job hunting only.

On the server the panel is a sandboxed systemd service with memory and CPU limits. A health check every two minutes restarts the service if it fails to answer three times in a row and reports it to the bot. Installation can be left to Claude Code: the /setup command installs the dependencies, builds the interface, starts the panel and walks the user through the steps only they can do — hh.ru keys, sign-⁠in, CV, bot.

An update waits for the panel’s jobs to finish, copies the database, switches the code and checks /api/health. Only a commit merged into master with green CI on Windows and Ubuntu goes out: 744 automated tests (625 for the backend, with mocks of hh.ru and Telegram, and 119 for the interface), ruff and strict mypy. Restoring from a backup is tested: a script brings the copy up as a second panel without hh.ru sign-⁠in or the bot and compares the data.

The browser sign-⁠in link lasts five minutes, and the bot reports the time, browser and IP of every sign-⁠in; without one, the public address serves only the sign-⁠in page. Only Caddy faces the internet, and it is locked down: binding the port is the only privilege it keeps, the file system is read-⁠only, and system calls are filtered.

The interface is built on its own design system: OKLCH colours, six text sizes, every component in every state on the /ui page. The page never scrolls sideways, tables scroll inside their own block, and every page is checked at 1280–⁠1536 px and on a phone.