Roles
How CareerBot sorts each open role to your directions, then ranks it, and what the level and score mean.
CareerBot reads every open role at your Target and Maybe companies from their own job boards: each morning, and when you choose Check roles again in Pursuits. A role gone from a complete read of its board is closed. Then each new role goes through two AI steps: a quick sort, then a full ranking. The point is to tell you which roles deserve your time, not to predict whether you'll be hired.
The steps
Deciding which roles deserve your time.
Sort roles
Once the descriptions of newly found roles are all in.
Reads
- Approved onlyYour directions: their names, positioning and titles
- PublicEach role’s title
- PublicEach role’s description, without what its company repeats in every posting
- PublicThe company’s name
May
Decide which of your directions a role could be for, leaning toward including it.
May not
Judge how well it fits. That’s the next step.
Rank roles
Right after sorting.
Reads
- Approved onlyYour projects and their facts
- Approved onlyYour directions and limits
- PublicEach role, and the directions the sort passed it to
- PublicEach role’s whole description, without what its company repeats in every posting
- PublicThe company’s name and one-line summary
- Approved onlyYour roles and their facts, and your years of work
- Whether to count how big a stretch a role is
- Rejected, to skipRoles you turned down, and why, as preferences, never as limits
May
Rate the role for each direction as strong, some, weak or none, with a score from 0 to 100 and a reason in plain words.
May not
Predict whether you’ll be hired. The score says how much of your time the role deserves.
Sort, then rank
Ranking reads a role in full against a direction, so it costs more than a glance. Sorting comes first to keep that cost down.
Sort roles decides which of your approved directions each role could plausibly be for, or none. It leans toward including: a role wrongly left out is never read again, while an extra one only costs a ranking. Only the directions a role is sorted to are ranked, and a role sorted to none isn't ranked at all. Those roles stay in the list, under Other roles.
Sorting waits until every new role's description is in. It reads each description without the text a company repeats across its postings (the same benefits and company blurb on every role), so the role's own work stands out.
The Sorting roles task's model sorts them, chosen in Settings, AI, under Advanced: the main model, or a model of its own. A chat model names the directions each role could be. A decision model (Jev, Liquid's d1 and others) answers for each direction whether the role's main work is that direction's work, with a probability. See Models and AI tasks.
Rank roles then reads each role against the directions it was sorted to, from the whole description, with what its board states and your approved directions, limits, roles, facts and projects. For each direction it gives a level, a score and one plain sentence of reason. When a posting is too thin to judge, the reason starts with "Read this one yourself:" and says what's missing.
Levels
| Level | What it means | Score band |
|---|---|---|
| strong | The work this direction is about, not beneath your level | 75 to 100 |
| some | Plausible and worth a look | 50 to 74 |
| weak | Only partly this direction's work | 25 to 49 |
| none | Different work, or against a firm limit | 0 to 24 |
What the score means
The score, from 0 to 100, says how worth your time a role is. A role is listed under its best direction, best score first.
With Count how big a stretch a role is on (it is unless you turn it off, in the menu in Pursuits), the level still says only how well the work fits. The score starts inside its level's band, then drops for each gap between the role and you: more years asked than you've worked (counted from your approved roles' dates, career breaks left out), seniority well above yours, core requirements your record lacks, and preferences it misses. A big stretch can take a score below its band. Each role lists its stretch in short plain words. A stretch lowers a score; it never hides a role.
With it off, the score only places the role inside its level's band, and gaps in your background never lower it.
A firm limit makes a role none. Roles that state something against a firm eligibility limit, like a security clearance above the one you hold or no visa sponsorship when you need it, are set apart under Against your limits. So are roles outside a firm location limit: on-site or hybrid somewhere you won't work, or remote when you don't work remotely, with the reason ("On-site in San Francisco; you work hybrid from Denver or remote"). A role that doesn't say how or where it's done is kept.
What else it reads from a role
Ranking also writes, for each role:
- Its details as the description states them: pay, setup, locations, seniority, years asked, employment type, travel, clearance and visa. Never guessed. What the job board states in its own fields wins over what ranking read.
- A brief: the job in two or three neutral sentences from the posting, and why it fits you and what's thin, from your approved record only.
These are what the role filters in Pursuits work on.
Ranking again
Nothing is ranked again by itself. When you change an approved direction's title or summary, roles ranked before the change are marked, and Rank again ranks them for your directions as they are now. It reads no job boards, and shows what it costs before you start.
A role whose own input fails three times in a pass is set aside until the next pass. If this month's AI budget runs out, the pass pauses and picks up where it left off. See Budgets and costs.
Edit this page on GitHub