A new page turns recruiter email into a job radar sorted by companyA new page turns recruiter email into a job radar sorted by company

September 29, 2026 · A new admin page reads recruiter email straight from Gmail and groups every software role it finds by company, refreshed by a digest each morning at 7:30.A new page automatically sorts recruiter emails from Gmail by company, and a short summary message arrives every morning at 7:30.
last updated

There's a new page at /admin/jobs that turns recruiter email into a job radar: every software role in the mailbox, filed under the company it came from. The new CRM package, jobs/, reads that mailbox over read-only IMAP, sends each new or changed thread to a model against a fixed schema for the role, company, comp and remote/hybrid fit, and files every match under its company in a SQLite store. A thread only goes back to the model when its content or the prompt version has changed, so a rescan of yesterday's mail costs nothing.

The page itself groups everything by company: a fit chip for remote, hybrid-in-LA or unstated, the comp if the email gave one, links to the actual posting, and the underlying emails each role came from. It reaches the CRM package the same cautious way every cross-repo read does here: if the data is missing or too old, the page reports that as an outage rather than quietly showing "no roles". A LaunchAgent runs the scan and sends a digest at 7:30 every morning, with the reload-safe habit this fleet keeps for anything on a timer: it only fires on schedule, never again on a restart, so nobody gets the same digest twice.

The build then got tuned against what actually showed up in the inbox, the same day it shipped. The first pass swept in job-board alerts alongside recruiter mail; being told "actually, no job alerts" turned that into a deny-list of LinkedIn, Glassdoor, Levels.fyi and Indeed senders, so only a recruiter's own message reaches the model now. A company can be dismissed from the page without losing its history. It keeps filing new roles in the background, and restoring it shows everything that arrived while it was hidden. Roles that never gave a location stopped being dropped, since silence about location isn't the same as ruling somewhere out. And the page now orders companies by whoever emailed most recently, reading the newest threads first, so a scan that gets cut short still files what just came in rather than losing it to whatever was oldest in the queue.

That model call (one thread, one schema, one answer) used to run on a metered API key, and building a whole radar in a day is exactly the kind of thing that runs the meter up. So the same day, the default for a model call changed across five projects at once: jobs/ and CRM's existing mail classifier, plus chess-coach, vintage-map and playlist-maker, all now reach the shared client's default provider, which is a Claude subscription running through claude -p rather than a paid-per-request key. None of the five need ANTHROPIC_API_KEY to run their normal path any more. chess-coach kept one deliberate exception: its opt-in metered transport still exists behind CHESS_CLAUDE_TRANSPORT=api, now pinned to the anthropic provider by name rather than picked up implicitly, so nothing starts spending real money on an API call without that flag being set on purpose.

There's a new page that reads recruiter email and turns it into a job radar, sorted by company. A new program checks the inbox where recruiter mail arrives (read-only, so it can look but never delete or file anything), and for every new or changed message about a job, it asks an AI model to pull out the company, the role, the pay if it's stated, and whether the job is remote, in-person in LA, or doesn't say. It only asks about a message once; if nothing about that message has changed since the last check, it skips it, so re-checking yesterday's mail is free.

The page itself lists every company with the roles found under it: a tag for remote, LA, or unstated, the pay if given, a link to the actual job posting, and the emails each role came from. If that data can't be reached or looks too old, the page says so plainly as an outage, rather than showing an empty list and letting it look like there's simply nothing new. A scheduled task checks the mailbox and sends a short summary message every morning at 7:30, and it's built so a restart of the machine never sends that same summary twice.

The page then got adjusted the same day, based on what actually showed up in the mail. It started out picking up automated job-board alerts as well as real messages from recruiters; being told "actually, no job alerts" turned that into a blocklist for LinkedIn, Glassdoor, Levels.fyi and Indeed's alert senders, so only an actual recruiter's message reaches the model now. A company can be hidden from the main list without losing anything about it. It keeps collecting new roles quietly in the background, and unhiding it shows everything that came in while it was out of sight. Roles that never stated a location stopped being thrown away, since not saying where a job is isn't the same as ruling every place out. And the list now shows whichever company emailed most recently first, checking the newest messages before older ones, so if a check gets interrupted partway through, it still has time to catch today's mail rather than getting stuck on something from weeks ago.

Asking the AI model to read a message and answer in that fixed format costs money each time, on a normal pay-per-use plan. And building this whole feature in one day, with every message it read, is exactly the kind of thing that runs that bill up. So on the same day, five different projects were switched to use a flat-rate Claude subscription for this kind of work instead of a pay-per-use key: the new job-mail reader, an existing tool that sorts other email, a chess-teaching tool, a vintage-map tool, and a music-organizing tool. None of the five need a pay-per-use key anymore for their everyday work. The chess tool kept one deliberate exception: it still has a setting to switch back to the pay-per-use option on purpose, and that setting was tightened so it can't be triggered by anything except someone actually choosing it.