Slushpile

Install

claude plugin marketplace add VonTerraProject501c3/slushpile

claude plugin install slushpile@slushpile

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

/slushpile:onboard

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.

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.

Materials quality: 8/10. The materials are not the problem.

Contrarian: SUBMIT_AS_PORTFOLIO_ONLY. Sending this cold spends an hour for a 2% shot. Two hours finding a referral is worth more than ten more cold applications.

One of these is a number about your document. The other is a decision about your afternoon.

The pipeline

The main pipeline

Three commands, in order. A search runs on these alone.

/slushpile:onboard              once per workspace — builds your profile,
                                preferences, and stories

/slushpile:job-board-search     search a careers board, extract postings,
                                score pool-anchored fit, contrarian gate,
                                create role folders

/slushpile:application-builder  build the resume and cover letter, then
                                iterate them against the review until they
                                stabilize

The three dispatched for you

/slushpile:application-builder dispatches adversarial-review, explore-experience and removing-ai-tells; /slushpile:outreach dispatches removing-ai-tells. Run one directly only to work on materials this pipeline did not build — a resume written elsewhere, a letter drafted by hand.

/slushpile:explore-experience   interview to surface experience you have
                                but never wrote down

/slushpile:adversarial-review   seven agents, five in parallel, verdict
                                per channel

/slushpile:removing-ai-tells    strip AI-authorship signals from prose,
                                with a gatekeeper on every change

Any time

/slushpile:outreach             find who they already know at the company,
                                grade the path, and draft the ask

/slushpile:redesign-templates   restyle the resume and letter templates,
                                holding the ATS constraints fixed

/slushpile:status               the queue, what is waiting on you, and whether
                                the pipeline's predictions are holding up

/slushpile:help                 what to run next, and how to read the output

The seven reviewers

The first five run in parallel and cannot see each other's work. The hiring manager sees all five. The contrarian sees everything, including the hiring manager, and can overrule it.

The manual

The rest of it lives in docs/:

• Getting started: what to gather before onboarding, and what to install.
• Skills: every command, and when to run it.
• The workspace: the files this writes into your directory.
• Your voice agent, Troubleshooting.
• Architecture: the diagrams, why the review is shaped this way, and how scoring and calibration work.

Your cover letters need your voice

An eighth agent writes the cover letter. It writes in one specific person's style, built from a corpus of their own writing.

slushpile ships aaddrick-voice as a working example so the pipeline runs out of the box. It is the plugin author's voice, not yours. Letters written with it will sound like a specific stranger — fine for seeing the pipeline work, wrong for anything you actually send.

Generate your own with written-voice-replication. It analyzes a corpus of your writing across 25 dimensions and outputs a voice agent, a voice skill, and a numeric profile with measurable targets. aaddrick-voice is that pipeline's own worked example.

Then point preferences.yaml at it:

voice:
  agent: "your-name-voice"
  is_mine: true

While is_mine is false, every skill that drafts prose warns you before it runs. That warning is the only thing standing between you and twelve applications sent in a stranger's voice.

Your data stays yours

/slushpile:onboard writes three files into your directory: profile.md, preferences.yaml, and stories.md. Every personal fact the pipeline uses lives there. Nothing is hardcoded into the plugin, and the repository has a CI gate that fails if a personal fact leaks into a skill.

That workspace will contain your full employment history, your compensation figures, and your constraints. Keep it in a private repository, or in no repository at all. The onboarding skill will say this to you and it will not initialize one for you.

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

Honesty is the feature

This tool will tell you that a differentiator you are proud of is median. It will tell you a role you want has a 2% conversion rate. It will occasionally tell you not to apply.

That is the product. A pipeline that grades most applications INTERVIEW and converts 5% of them is not producing signal, it is producing optimism, and it will keep doing that indefinitely because nothing in it ever pushes back. The contrarian pass, the pool anchoring, and the per-channel probabilities all exist to make the output usable specifically when it says no.

There is a Calibration table in the workspace tracker for exactly this reason: you record what the pipeline predicted and what actually happened, and the priors get corrected by your own history rather than by anyone's confidence.

Tune it

The ten skills and eight agents are Markdown. Fork, edit, install your copy:

claude plugin uninstall slushpile

claude plugin marketplace remove slushpile

claude plugin marketplace add <your-username>/slushpile

claude plugin install slushpile@slushpile

If you change a skill, run the gates before you push. See CONTRIBUTING.md.

License

MIT. See LICENSE.