slushpile

Slushpile Seven agents try to reject you before a recruiter gets the chance. What they find, you keep.

Connect on LinkedIn!

English · 简体中文 · Español · Português (BR) · Tiếng Việt · AI Bro

Install

Claude Code

claude plugin marketplace add VonTerraProject501c3/slushpile

claude plugin install slushpile@slushpile

Then, in the directory where you keep your job search:

/slushpile:onboard

Codex

codex plugin marketplace add VonTerraProject501c3/slushpile --ref main

codex plugin add slushpile@slushpile

Codex prefixes plugin skills with the plugin name:

$slushpile:onboard

Codex has no subagent dispatch, so the review pipeline runs its seven personas sequentially in one context instead of five of them in parallel. Same output, slower, and slightly more prone to one persona's reasoning bleeding into the next.

Cursor, Gemini CLI, and manual install

See INSTALL.md.

The problem

You are not being graded against the job description. You are being graded against the other seventy people who applied to the same requisition this week.

Almost every tool in this space gets that backwards. Feed a resume and a posting into a resume optimizer and it will tell you your keyword match went from 68% to 91%, which is a real number about the wrong question. If the 75th-percentile applicant in that queue matches at 94%, your 91% is a rejection, and nothing in the tool will ever tell you so.

The second thing they get wrong: they return one verdict. But the same resume and the same letter convert at maybe 2% through a cold portal submission and 30% through a referral. Those are not the same decision, and collapsing them into a single "strong match" is not a simplification. It is an error with a confident interface on top of it.

The third is the one nobody names. These tools have no memory. You paste a resume in, you get a number back, you close the tab, and the tool ends the session knowing exactly what it knew at the start. A search is forty applications over three months. Every one of them costs full price.

What this does instead

It builds a model of you, once. /slushpile:onboard interviews you and writes three files: a profile, a preferences file, and a set of stories. The profile is not a resume — it is the pool a resume gets cut from, several times longer than anything you would ever send. Every later stage reads it, and nothing asks you those questions twice.

It tries to talk you out of the role before you write anything. The search stage scores each posting against the estimated applicant pool, runs kill criteria, builds an expected-value matrix per application channel, and puts a contrarian in front of the tier list. Every other tool starts working after you have already decided to apply. The expensive mistake happens before that, and this is the only stage that can still catch it for free.

It goes and opens the channel it says is best. A matrix that prices a referral several times above a cold submission is worth nothing if nothing acts on it. /slushpile:outreach reads your own history for someone you already know there, asks you the one question no file can answer, researches named targets from public professional presence only when you know nobody, and drafts the ask in your voice. It grades each path by what that person could actually say about your work, then records it in job_search.md, where the next assessment reads whether that channel is open at all. It never sends anything.

It attacks what it just wrote. A model asked whether its own draft is any good will tell you yes, at length. So the builder does not ask. It hands the resume and the letter to seven reviewers, five of whom run in parallel and cannot see each other's findings, each given only what its role would really have — the eleven-second screener is never shown the cover letter, because a screener who read the letter is not one. The builder fixes what comes back and sends it through again. Round two has to hold before it will let you submit, and it stops at three rounds, because past that the gaps are structural and more editing is motion.

It writes what it learns back into you. When a review says a section is thin, the interview that follows usually finds the experience was real and you never wrote it down. That goes into the profile permanently. The conversion estimates get corrected against the outcomes you record. Your twentieth application starts from a better place than your first, which in every other tool is simply not true.

What comes out is a verdict per channel — cold submission, warm referral, cold outreach, inbound from your public work — each with a probability range rather than a verdict word, and with materials quality scored separately from expected value. "1-3% interview" is information. "MAYBE" is not. Excellent materials sent to a wrong-fit role still have low expected value, and those two numbers routinely disagree.

What changes

Same resume, same posting, same afternoon.

What a resume optimizer tells you

Match score: 91% ✅

Great news — your resume is a strong match for this role!

✅ 14 of 16 required keywords found ✅ ATS-friendly formatting detected ⚠️ Consider adding: "stakeholder alignment", "OKRs"

You're ready to apply!

What Slushpile tells you

Pool position: p55. Median applicant here has shipped at comparable scale at a company the hiring manager recognizes. Your open-source work is real and it is not rare in this pool — roughly a third of the p75 cohort has something equivalent.

Cold submission: REJECT, 1-3%. The years-of-experience dropdown on the form gates at 8. You have 6 in the titled function.

Warm referral: MAYBE, 20-30%. This is the only channel with a real path.

**Material