Slushpile 1.0.0 — an adversarial job search with a memory

Slushpile grades your resume against the other people who applied to the same requisition that week, not against the posting's keywords.

It's a job application pipeline that runs inside your coding agent. Nine skills and eight agent definitions, all Markdown. It builds the resume and cover letter for a posting, then spends most of its effort trying to talk you out of sending them.

Seven reviewers, and they argue. Five run at once and can't see each other's findings. A hiring manager then reads all five. A contrarian reads everything after that, including the hiring manager, and can overrule it.

A verdict per channel, with a probability range. Cold portal submission, warm referral, cold outreach, and two kinds of inbound each get their own number instead of one blended answer. Materials quality is scored separately from submission expected value, and those two routinely disagree.

It remembers. Onboarding writes profile.md, preferences.yaml, and stories.md into your own directory. What the review finds gets written back, so later applications start from a better place than the first.

It never submits anything. No skill touches an application portal, an email, or a form. It writes files and you send them.

Cover letters come from a voice agent built on a corpus of your own writing. One ships as a working example so the pipeline runs out of the box. That one is my voice, and there's a separate tool for generating yours.

Nothing in the pipeline hardcodes a personal fact. Every one of them lives in your preferences.yaml or profile.md, and a CI gate fails the build if one leaks into a shipped skill.

Before you install

It's built to be usable when it says no. It will tell you a differentiator you're proud of is median, and it will sometimes tell you not to apply.

Claude Code is the only supported harness with subagent dispatch, so the five parallel reviewers genuinely run in parallel there. On Codex, Cursor, and Gemini CLI the seven personas run sequentially in one context. Same output shape, slower, and slightly more prone to one reviewer's reasoning bleeding into the next.

Install pdftotext if you can, so the reviewers read what an applicant tracking system reads rather than what a PDF viewer shows. A LaTeX toolchain is optional and only matters if you use the two shipped document templates. Everything else works on extracted text.

Install

# Claude Code
claude plugin marketplace add aaddrick/slushpile
claude plugin install slushpile@slushpile

# Codex
codex plugin marketplace add aaddrick/slushpile --ref main
codex plugin add slushpile@slushpile

# Gemini CLI
gemini extensions install https://github.com/aaddrick/slushpile

Then run /slushpile:onboard in the directory where your job search should live. Don't use a checkout of the plugin for that. The workspace is going to hold your employment history and your compensation figures.

MIT licensed. Cursor and any other harness are covered in INSTALL.md, and README.md walks through how the pipeline works.