The interview after the research

A 30-minute phone screen usually goes one of two bad ways. Either you wing it off the resume and ask the same five generic questions you ask everyone, or you over-prepare a script and spend the call reading it instead of listening. Both fail the same way: you hang up, realize you forgot to probe the one thing that actually decides it, and write up a verdict that’s mostly vibes.

I wrote earlier about running candidate background research with parallel Claude Code agents. ¹ That workflow told me who was real and ranked the ones who were. It stops exactly where this problem starts. The research picks who I talk to. It does nothing for the 30 minutes once the call connects. So I built the same thing one stage later, on the same bet: the reusable artifact is what compounds, and the model just renders it.

None of the scorecard half of this is new. Structured interviews have beaten gut-feel hiring on validity since the 1990s, and scoring against fixed, job-relevant dimensions is the most replicated finding in the field. ² The new part is narrow: an agent now preps and grades that structure for me, per candidate, from research it already did.

#The artifact is the interview page

For every candidate who advances, a parallel agent generates an Initial Interview page straight from that candidate’s background-research page. One wave produced twenty-two of them in an afternoon. The page is not notes I take during the call. It’s the call, planned in advance, structured so the writeup is half-done before I dial.

The trick is the same one that made the research workflow scale: hold the scaffolding constant, vary only the tailored part. Most of the page is byte-identical across every candidate. A small, specific slice is theirs alone.

The constant scaffolding is three things: a loose four-beat arc, a six-dimension scorecard, and a talking-points cheat sheet collapsed behind a toggle. The arc is just open and “why this company,” then their strongest shipped work, then the central probe, then next steps. The scorecard is the part that matters, and it’s the same six dimensions, scored one to four, every call:

Dimension	A 1 sounds like	A 4 sounds like
Shipped craft	Can’t navigate their own code	Walks the hard parts of something real they shipped
Domain depth	Generic answers that fit any platform	Names specific failure modes and how they handled them
Systems thinking	Feature by feature	Reasons about tradeoffs, scale, and what breaks first
Product judgment	Built what the ticket said	Owns what to cut, and why
Communication	Waits to be asked	Thinks out loud, gets to “I don’t know” fast
AI fluency, lived	Lists the tools	A real workflow plus where it fails them

The scorecard being fixed is the point that’s easy to miss. It’s not there to make me a better interviewer. It’s there so I can compare twenty-two candidates against one rubric instead of against twenty-two different moods on twenty-two different afternoons. Being a sharper interviewer doesn’t buy you that. Only the fixed rubric does.

#The probe bank does the work

The tailored slice is a small table the agent builds from the research: the question, the scorecard dimension it feeds, and the binary thing I’m listening for. Six to eight rows, one flagged make-or-break. Three of them from a real page, lightly redacted:

Probe	Dimension	Listen for (advance / concern)
Make-or-break. Walk me through the playback bug that cost you the most time, and how you found it.	Domain depth	A specific failure mode and how they isolated it / a generic answer that fits any bug on any platform
Tell me about a decision you reversed after shipping. What changed your mind?	Product judgment	Owns the call and names what the data said / blames the spec or the deadline
Show me something real you’ve built with an AI coding tool that you’d defend.	AI fluency, lived	A workflow plus where the tool fails them / a list of tool names with no opinions behind them

That last row is how a 30-minute call tests lived AI fluency instead of the keyword: ask for one artifact they’d defend, and listen for whether they can say where the tool falls down. The vocabulary is free. The opinion isn’t.

The probe bank isn’t a questionnaire I march through top to bottom. It’s a priority list for a conversation that goes wherever the candidate leads. If they go deep on something unprompted, I follow. But I always know which questions still need answers before I hang up, and which single one I can’t leave without. The make-or-break flag is there because most screens have exactly one question that decides the outcome, and the common failure is not getting to it until the call is already over.

The probe bank isn’t a script to read. It’s a map of what I can’t hang up without knowing, with the make-or-break question flagged on top.

#The page grades itself

Here’s the part that pays back the setup. Because every probe is tagged to a scorecard dimension before the call, the call can grade itself afterward.

A probe bank, where each probe is pre-tagged to a scorecard dimension, feeds the interview. The interview produces a transcript. An AI pass reads the transcript against the probes and fills six dimensions, each scored one to four with evidence, then drafts a recommendation. The human decides.

The conversation gets transcribed onto the same page (the notetaker handles that automatically). An AI pass reads the transcript against the probe bank, fills each of the six dimensions with a one-to-four rating plus the evidence from what was actually said, drafts a recommendation, and writes the one-line entry for my activity log. Then it stops. Nothing changes the candidate’s status, nothing commits a score, nothing sends. It hands me a graded draft and waits.

That stop is the whole design. Mapping what they said to the dimension it speaks to, summarizing the evidence, drafting the writeup: that’s work an AI does well and I do tediously. Deciding whether it’s a yes stays a button I press. This grading pass is the newest piece of the system and the one I trust least on its own, so I read every transcript myself and override the draft rating whenever it missed the room. The draft saves me the writing, not the judging. What makes it worth trusting at all is the tagging: I can see exactly which answer drove which score, because the probe pointed at the dimension before I ever asked the question.

#Grading a transcript beats grading a memory

The obvious objection is that an algorithm scoring a job interview is exactly the thing you should be nervous about. It is, when it’s a black box rating a person. This isn’t that.

The pass doesn’t score the candidate. It scores the transcript against the job-relevant probes the page declared before the call, the ones the agent drafted and I approved when I prepped. Every rating points back at the specific thing that was said. That’s the opposite of the failure mode it replaces: me, three calls later, scoring a candidate from memory and a feeling. A rating I can trace to a quoted answer and a pre-declared probe is more auditable than gut feel, not less, and the audit trail is exactly what you want if a hiring decision is ever questioned. The human still decides. The machine just makes the decision show its work.

#What it doesn’t do

Same discipline as the first post. This does not decide who to hire. It removes the tax that made screens bad, the prep nobody does and the after-the-fact scoring nobody trusts, so the 30 minutes of actual conversation gets all the attention. A candidate who surprises me in either direction overrides anything the page predicted, which is why the make-or-break probe exists: so the surprise has somewhere to land.

And it doesn’t fix a bad interviewer. A weak conversation with a great prep page is still a weak conversation. What the page buys is that I show up to every call having done the thinking, asking the questions that matter for this specific person, grading all of them on one scale. The research told me who was worth 30 minutes. This is what makes the 30 minutes worth taking.

Notes

1. The first part, Running candidate background research with parallel Claude Code agents, covers the research workflow this one picks up from. Part one decides who to talk to; this one is how to talk to them. ↩
2. Frank Schmidt and John Hunter, "The Validity and Utility of Selection Methods in Personnel Psychology" (1998), is the canonical meta-analysis: structured interviews and work-sample tests sit near the top for predictive validity, unstructured interviews well below. The probe-bank-tagged-to-dimensions structure is that finding operationalized so an agent can prep and grade it. ↩