How jobwhat learns

jobwhat isn't a static set of rules that ships once and stays put. It gets better at filling your forms the more you use it, and it gets better at filling everyone's forms as more people use it — at three different levels, each with a different answer to "who taught it that."

flowchart LR
    subgraph you [Yours alone — stays in your browser]
        L1["Level 1\nCustom Q&A you write"]
        L2["Level 2\nAnswer-and-remember\n(auto-adds to Custom Q&A)"]
        L2 --> L1
    end
    subgraph everyone [Everyone's forms — label & options only, never an answer]
        L3a["Level 3 · per-job\nShare a stuck question's shape\nwith the next applicant"]
        L3b["Level 3 · aggregate\nDaily-reviewed synonyms\nacross every board"]
    end
    L2 -. required question\nnothing could answer .-> L3a
    L3b -. sharper matching .-> L1
    L3b -. real examples .-> AI["AI answers model"]

Level 1 — you teach it directly

Custom Q&A (Workday settings, but used everywhere) is a keyword → answer pair you write yourself: "largest number of engineers" → 10+. The keyword just has to appear somewhere in the question; the answer is matched against whatever the real form offers, so one entry works across dropdowns, radios, button groups and plain text. See Greenhouse, Ashby & Lever → Custom questions.

This is the fastest way to fix something specific to your answers — a question only your industry asks, a preference no built-in rule could guess.

Level 2 — it teaches itself, from you, automatically

Answer-and-remember closes the gap for required questions that stump everything — no rule, no custom Q&A, no AI. The badge asks you once, fills the field, and quietly appends the answer to your custom Q&A under the question's own content words. The next form — any board, any company — that asks something containing those same words is filled without asking again.

No setup, no review: you only ever answer a given required question once, and jobwhat remembers what "asking again" would have looked like.

Level 3 — it learns from everyone else

This is the level where jobwhat gets smarter independent of anything you personally typed.

Per-job, immediately. When a fill leaves required questions nothing could answer, jobwhat can share just the question's label and options — never your answer — under a fingerprint of that exact posting. If someone else queues the identical job next, they're asked up front, before their run even starts, instead of hitting the same wall. Your Metrics tab shows both directions of this as Questions you shared and Heads-ups from others' forms. Off switch: Settings → Share the questions a form asked that nothing could answer.

Across every form, over time. Separately, jobwhat keeps an aggregate count of which question wordings keep showing up as unanswerable on each job board — again, only the label and a sample of its options, never an answer, and never tied to you. A person reviews the frequent ones periodically and writes a small set of synonyms: "this wording usually means the yearsExperience field." That reviewed list ships as a daily-refreshed overlay that:

• lets the built-in matching rules recognize an oddly-worded question they'd otherwise miss, and
• gives the AI answers model a handful of real examples of how similarly-worded questions map to profile fields, so it guesses better on the ones it does see.

It only ever fills a gap — a reviewed, shipped rule always wins if one already matches, so this can't override or surprise you on something jobwhat already handles.

The Get set up card in the popup is where this loop reaches back to you: if forms keep asking something neither your profile nor your custom Q&A covers, it's listed there — "N questions forms keep asking that your profile can't answer" — with an inline box to answer it once. Answer it, and it's saved as custom Q&A like anything else, closing the loop back to Level 1.

What never happens

At every level, what travels is the shape of a question — its label, its options — never an answer. Your own answers stay in your browser (Levels 1–2) or go only to the AI provider for the one question you opted into sending (see exactly what gets sent). Nothing here trains a model on your data, and nothing here is personalized to you except what you typed yourself.