Agent skills for screening resumes
A screening skill reads a stack of resumes against criteria you wrote down, and returns a short structured note per candidate: what they have, what is missing, and the one question you should ask on the call. It sorts and summarises. It does not decide, and it should never be the thing that rejects a person.
Write the criteria before you read anyone
The value is almost entirely in this step, and it is the step people skip. Criteria written after you have read four resumes are a description of those four resumes.
Must-haves, stated as evidence.
Not "strong SQL" but "has shipped something that queries a production database, named in the resume". Evidence you could point at is checkable. An adjective is not.
Nice-to-haves, ranked.
Three at most, in order. An unranked list of eight nice-to-haves produces a ranking driven by whichever candidate happened to write the most.
Explicit non-signals.
Name the things you do not want weighed: school, employer prestige, gaps, tenure length if you have decided it does not matter. Saying so in the criteria file keeps them out of the summary.
The output shape.
Fixed fields, same order, every candidate. A free-form paragraph per person is unreadable at forty resumes and impossible to compare.
What a screening pass returns
> screen the 38 resumes in inbound/backend-role against
criteria/backend-mid.mdCandidate: A. Okafor
Must-haves: 3 of 4
met production database work (payments service, 4 yrs)
met owned an on-call rotation
met wrote the migration tooling, named in resume
missing no evidence of working on a team under 10
Ranked nice-to-haves: 2 of 3 (typed language, infra as code)
Ask on the call: the payments service is described as
"rebuilt" twice. Which parts did they personally write?
Not assessed: school, employer, gap 2022 to 2023The last two lines are the ones that earn the skill its place. A named question per candidate makes a first call twice as useful, and an explicit list of what was not assessed lets you check that the criteria held.
Where the line is
- A skill does not reject anyone. It sorts a stack into an order and writes notes. A human reads the bottom of the stack too.
- A summary is not evidence. If a claim matters, it is checked against the resume itself, not against the note the skill wrote about it.
- Screening carries legal and fairness obligations that vary by where you hire. A skill inherits none of that judgement, and writing the criteria down does not discharge it.
- It cannot tell you whether someone is good. It can tell you whether the page in front of it contains the evidence you asked for.
Keep the criteria file and the skill in the same commit
Criteria drift, roles change, and a screening skill pointed at last quarter's criteria file will produce confident, consistent, wrong answers with no error message. Keep the criteria beside the skill, date them, and re-run a handful of resumes you already judged by hand whenever either one changes. That is the whole of the maintenance, and there is a short guide on testing whether a skill actually works.
Works with Claude Code.
The Recruiter pack
Twelve skills for screening, outreach drafts, interview notes and pipeline reviews, with the criteria templates and the eval cases included so you can see what was tested. The pack page carries the price, the refund terms, and the date it was last tested.
See the packs