Claude Cut A Couple Hundred Profiles To 20. Then He Rebuilt The List By Hand.

First thing he asked me on the call was what time it was where I was. It was 8:30 in the morning for me, and he said that’s not too bad. Then we got into it, and somewhere in the middle of walking me through how he builds a candidate shortlist he stopped and asked me:

“Do you have a sense of how we can use AI to make that less effort?”

That was Friday, October 2. He and his partner run a small recruiting and staffing firm, it’s the two of them and one or two other people, and they’d only recently started doing exec search. Earlier on he’d already told me “right now we use AI a lot, but probably very inefficiently.” I liked that he just said it like that.

So this issue is basically my answer, a few weeks late and a lot longer than what I gave him on the call. He isn’t a client and I haven’t built anything for his firm, the call ended with me owing him a proposal. So what you’re getting is what he told me, quoted from the recording, and what I would test. No results in here, because there aren’t any yet.

How He Builds A Shortlist

He writes the job description and turns it into a rubric. Then he goes into LinkedIn Recruiter and searches by location and title and keyword, and he keeps changing the search so it stays fresh, and keeps adding people to a pipeline until he has, in his words, “call it 200 people.” LinkedIn lets you download the profiles in bulk so he does that, and then he’ll “throw that in Claude with the rubric” and ask which ones are worth reaching out to.

I would probably do the same thing to be honest. He called LinkedIn’s filters “very brittle,” and what he’s actually screening for doesn’t sit in any filter:

“So some people write zero to one on their LinkedIn. But most people don’t. They just say first employee or 20th employee, and that’s like a proxy.”

Vertical SaaS was another example he gave. You look at a logo on someone’s profile and you have no idea what that company sells, somebody has to go and look it up. Claude was doing that reading for him and he rated it, “very powerful for this” was how he put it. And then he got to this part:

“And so Claude will take it from a couple hundred profiles down to 20. And then we manually make a new pipeline that’s just like the curated 20.”

After that they message those 20 people, track who replies and follow up, all by hand as well. Out of the whole thing, outreach included, this is the step he picked as the slow one:

“I think the part that’s like the most, you know, time-consuming is building the pipeline.”

Later in the call it got a bit messier. Claude isn’t on the same account as everything else, so the output goes into a different Google Sheet and he moves it around himself. “Drowning in CSVs” was his phrase. He likes Google Sheets and said he’s good at them, and then straight after, “it’s definitely pretty manual, right?”

I know that one. At my old job I was on Excel sheets every day, mostly on my own, so yeah.

So the AI step was actually the one part he was happy with, and the time was going everywhere around it. Downloading the profiles, pasting them into Claude, reading what comes back, rebuilding the pipeline by hand, moving output from one sheet to another, and every one of those is a person sitting there doing it. If your AI step still needs someone to download, upload and paste on both sides of it then I wouldn’t call that automated yet. It’s a faster copy machine with a person walking the paper between rooms.

Two Robots

The easy reading of that call is “build them a CRM.” He pushed back on that himself, saying “we don’t need to go crazy building this whole CRM.” A clean, written process would do. The thing he actually wanted to figure out, from his side, was:

“What should be AI. What should be a recruiter?”

Then he told me about his partner. His partner had sent him a pipeline, and he was pretty sure the partner built it with Claude. He ran it through his own Claude, and his Claude came back saying the candidates were a mix of too junior and too senior, and he agreed with it. His summary was “we’re just having our two robots.”

What I said back to him was that “this is AI versus AI versus AI,” and then a bit later that once AI is scoring the AI and AI is applying, “it does become a bit of a mess.” I still think that’s where a lot of recruiting is right now, he’s obviously seen way more of it than me though.

If you’re cleaning up a process like his, keep two things apart. A recruiter deciding if someone fits the role is judgment, you want to keep that. Moving someone who’s already been reviewed from one list into another list is just carrying. They happen right next to each other, so people tend to time them together, and then nobody can tell which one is eating the afternoon. So I’d put them on separate lines.

I’d also be careful with the 200 to 20. That’s how he described the funnel, nobody on the call checked how good Claude’s picks were, and a shorter list isn’t automatically a better one. When I sketch a screening step for a firm like his, I want Claude to give a reason for every person it keeps, pointing back to something that’s actually on the profile, and to flag the ones where the evidence is thin so a recruiter goes and looks at them properly instead of trusting a confident guess.

The Copy Paste Audit

This is the worksheet I’d use on your next shortlist build. I laid it out from the steps he described and I haven’t run it across a group of firms yet, so treat it as the first thing I’d want to see from you rather than some tested method.

For each step, write down what moves, where it moves to, who moves it, whether a person does it by hand and how many minutes it took. Use the real destination, account included, because “it’s in Sheets” hides exactly the kind of problem he had.

1. Search. Search results go into a candidate pipeline. Who ran it, and how many minutes?

2. Download. Profiles go from LinkedIn into files. Was it one bulk export or a few?

3. Paste. Profiles and the rubric go into Claude, or whatever you screen with. How many batches and how many minutes?

4. Screen. Claude’s answer becomes a reviewed selection. Put the review time on its own line, because this is the judgment you want to keep.

5. Shortlist. Reviewed candidates go into the pipeline the recruiter will actually do outreach from. Is that rebuilt by hand?

6. Extra transfers. Output moving between sheets, accounts or tools. Give each move its own line.

Yours probably won’t look exactly like his. If someone exports a file, fixes the columns and uploads it again, that’s a separate line. If a step genuinely happens without anyone touching it, write that down too. Just don’t mark a whole step as automatic because Claude is somewhere in the middle of it.

Once it’s filled in, pick one handoff to look at first. For his workflow I’d start with the move from Claude’s selection to the curated pipeline, because he told me outright he rebuilds it by hand. Without his timings or a look at what his accounts allow, I can’t promise that’s the biggest win, or that it goes away completely. He brought up LinkedIn being a “pretty, like, closed system” himself, and being able to download profiles doesn’t mean every step after that can be connected. Where it can’t, I’d write down the manual step that’s left and stop calling it something the system does.

Start With One

Full confession on this part. Earlier in the same call he’d said they wanted “AI like all over the business,” and I went for it. I pitched what I called an AI operating system, one in-house dashboard with different sidebars for the scraping, the job postings, the outreach, tracking the pipeline, everything on one screen. He said that was fair enough but where would I actually start? And I walked it back:

“I initially start off with running these set systems or workflows for, like, for example, just one system first, and we can prioritize.”

That’s what I’d tell any firm reading this as well (and yes, I should have led with it). The shortlist build is one workflow. It starts with a role and a rubric, and it’s done when there’s a reviewed shortlist sitting where the recruiter actually works from, with every person on it traceable back to their profile. Prove a change there against a real role, with a recruiter checking the output before anyone gets contacted. Then compare the whole build against your baseline, including the time spent checking and fixing the output, because a faster screening step doesn’t help much if someone then spends longer repairing the list.

Do This Today

Time your next shortlist build, from the first search to the moment the reviewed shortlist is ready where your recruiter will use it. Write the minutes next to each step of the audit while you’re doing it, I’d much rather work from real minutes than a guess you make afterwards.

Keep the start and finish times as well as the active minutes. If Claude is processing while you go do something else, note that as waiting so the same stretch doesn’t get counted twice. At the bottom write the one handoff you’d fix first and what you’d need to see to call it fixed. Keep the sheet with the shortlist and save a blank copy for the next role, so you’ve got something to compare against next time.

And if you know a recruiter who’s still pasting profiles into Claude one batch at a time, send this to them.

If you subscribe, reply to this email with “kit” and I’ll send you the AI Ops Starter Kit. It’s five n8n workflows for catching enquiries that go unanswered and automations that fail quietly, with sample data, a setup guide and the Enquiry Leak Checklist. All five passed my own test runs. Nobody outside Robossist has set it up yet though, so I’m not claiming any time saved.

One more ask before you go, and it doesn’t have to be about recruiting: reply with the step where your enquiries go quiet and I’ll break down three of them in a later issue.

Dhruv