People
Application summaries, not candidate scores
A worker reads every application in the same shape against the criteria you published. It does not rank anyone, and it does not reject anyone.
What the worker does
An AI worker in Polaris reads applications from a connected Gmail inbox and Google Drive and returns one structured summary per candidate against the criteria written for the role: what evidence exists for each requirement, what is missing, and the exact quote it came from. It produces no score, no ranking and no rejection. A hiring manager reads the summaries and decides who to interview.
- Connections
- Gmail · Google Drive
- Produces
- One evidence summary per candidate
- Does not produce
- Scores, rankings, rejections
The reason screening goes wrong is that it is done inconsistently, late at night
Two hundred applications arrive. The first thirty get read carefully. The next hundred get eight seconds each. By the last seventy the reader has an unspoken heuristic that has nothing to do with the job description, and nobody can reconstruct why anyone was cut.
Automated scoring makes that worse rather than better, because it takes an unexamined judgment and gives it a number, which makes it look defensible.
The useful intervention is consistency of reading, not automation of deciding. A worker reads all two hundred with the same attention and the same questions, quotes what it found, and says plainly where the application does not answer a requirement. Then a person makes every call, with the evidence in front of them.
What a summary contains and what it deliberately omits
In every summary
- Each published requirement, with the evidence quoted verbatim
- Requirements with no evidence in the application, named as such
- Dates and durations as the candidate stated them
- Anything the application says that does not fit the stated criteria but is factually notable
- A link back to the source document
Never in a summary
- A score or a percentage match
- A rank against other candidates
- A yes or no recommendation
- Inference about age, nationality, health, family or any protected characteristic
- Anything sourced from outside the application unless you explicitly asked for it
How to brief the worker so the output is fair
The quality of the screening is set by the criteria, and the criteria are yours.
Write the criteria before the first application arrives
Put them in the SKILL.md. Criteria invented halfway through a pile are the mechanism by which bias enters a process.
Make each criterion evidence-shaped
"Has shipped a production system they were on call for" can be evidenced or not. "Strong engineer" cannot, and asking a worker to assess it produces noise dressed as a finding.
Instruct it to quote, not to characterize
A quote can be checked against the source in two seconds. A characterization cannot.
Name the human who decides
The task has an assignee. Make the shortlist a task assigned to a person, and keep it separate from the summarizing task.
Where the work goes
| Stage | Who | Output |
|---|---|---|
| Define criteria | Hiring manager | Criteria in the role's SKILL.md brief |
| Read applications | Worker | One structured summary per candidate |
| Flag gaps | Worker | Requirements with no evidence, named per candidate |
| Shortlist | Human, always | A named list with reasons |
| Reply to candidates | Human | Sent from a person's mailbox |
Questions people ask
+Can we ask it to rank the top ten?
You can ask, and you should not. A ranking from a worker gets treated as an assessment even when everyone agrees it is only a suggestion, and that is precisely the failure mode this page exists to avoid. Ask instead for the summaries sorted by application date and do the ranking yourself.
+Will it search the web for a candidate?
Only if you explicitly instruct it to, and that is worth thinking hard about before you do. Unsolicited background research on applicants creates fairness and data-protection problems that a hiring process does not need.
+What does a screening run cost?
It is billed on human-equivalent hours from the work log at roughly $2 per hour, computed from observable effort: documents read, prose produced, files generated. A large pile of applications is a real job and the log itemises it line by line so you can check the arithmetic.
+Can several people see the same summaries?
Yes. The delivery is a comment on a task in a shared workspace, so the whole hiring panel reads the same document rather than four people forming four impressions from four different reads.
Related
HR work an AI worker can prepare, and where it must stop
Scheduling, onboarding logistics, policy drafts and review-cycle admin, prepared by a worker. Every decision about a person stays with a person.
Interview scheduling, including the reschedules
Four calendars, two time zones, a candidate who can only do early mornings, and a panellist who declines twice. This is the job.
Onboarding that is finished before the first Monday
Nobody's first day should start with an apology about accounts. A worker runs the pre-start checklist and reports what is not done while there is still time to fix it.
Hire an AI recruiter
Every candidate is scored against the same written rubric, and the rubric is a file your hiring manager wrote and can change.
Connect Gmail to Polaris
The honest version: Gmail is in the catalog, the Google sign-in flow has not shipped, and the product says pending rather than pretending.
Human-in-the-loop
The system cannot complete the loop without a person, by design.
Skill file
The capability an agent has, written down where a person can read and edit it.
Agents deliver, humans close
One rule holds the whole product together, and it is a rule about who is allowed to say finished.