Tailor job applications with subagents that reject unsupported claims

Shrivastava set up a team of seven Claude Code subagents behind one /apply command. Given a master resume and a job link, a job analyst splits the posting into must-haves and nice-to-haves, a researcher collects company facts with a quoted source for each, and a fit checker returns APPLY, STRETCH or SKIP. Only after an APPLY or STRETCH do the writers produce a tailored resume, a 150 to 250 word cover letter and interview answers, saved in a folder per job with a Word export. A truth checker on Opus, with read-only tools, compares every line against the original resume and sends drafts back. The writers have no web access, and a plain script checks every number. He tested it on an invented candidate against real postings and on twelve test cases he wrote.

For a Customer.io role the fit checker said SKIP because of location and language rules, after about 2 minutes and $0.32. For an Atomicwork role it said APPLY and produced the full pack in about 7 minutes for an estimated $1.79. The truth checker ran four rounds and rejected small overstatements, including an invented weakness. Three minor fixes were left unchecked at the round limit, so the pack was marked for a human read. The test set scored 11 of 12 at first; the miss was his own clean letter, which held unsupported claims, and then 12 of 12. He thought one APPLY should have been a STRETCH.

Nothing is sent anywhere; the person reads the pack and applies by hand. The limits are set through each subagent's allowed tools, not only the prompt. His first manager agent refused to start on a summarised posting and had to be loosened, since many job sites block automated reading. The candidate in his tests was made up.