The Parts the Machine Could Not Do

I spent a day rebuilding two automated job searches with Claude Code. I am in the middle of a career shift and looking for more challenges. Claude read error text, probed more than twenty sites, tested assumptions and drafted the fixes. The thing that actually broke the deadlock was a URL I pasted in from memory. The site I already knew worked, and it had a track record in my world of teaching.

This is the third of three posts about that day. The first two are Rebuilding a Job Search That Had Stopped Working, on what was broken, and A Successful Response Is Not a Successful Result, on how to test for it. This one is about the parts that did not yield to systematic probing, and about checking claims, including the confident ones.

The best result of the day came from a human

Partway through, I pasted in a search URL I use by hand, the one I actually open when I am looking for teaching work. It revealed that the board has a real, working, server side subject and radius search that the automation had never touched. It had been guessing keywords instead, badly, for months.

That one URL replaced five searches with one and made the whole rebuild possible.

The AI could test a hundred URLs in an afternoon. It could not know which one a working teacher actually opens. Twenty five years in an industry is a kind of index, and it is not one that any amount of systematic probing reconstructs.

Human in the loop

That is the honest shape of this kind of work, and it is worth saying plainly because most accounts of it are triumphant in a way the actual afternoon was not. The useful contributions from the AI were unglamorous: reading an error message carefully, and testing assumptions instead of accepting them. The judgement about what mattered stayed with me, and several of the calls that day went against what the tooling suggested.

Checking a second opinion, including the AI’s

I reviewed two sets of outside recommendations during the session. One was a technical review from a different AI model. One was a widely shared “thirty days to get hired” list from a social media influencer. Both were checked against evidence rather than accepted.

As an aside: run the same task past a second model when it matters. That is where I caught the four claims below.

What the other AI review got wrong

It contained genuinely useful suggestions, and it also produced four claims that did not survive contact with the system.

The last one is the serious one. Following it would have reinstated a bug that had already been found and paid for once. A confident recommendation, cleanly argued, pointing straight back at a decoy.

What the social media influencer got wrong

About half of it was sound, and that half matched conclusions my own research had already reached independently: build a target list, go direct to hiring managers, talk to people who started the role recently, and spend half your time on conversations rather than applications.

The other half was affiliate shaped. Three product placements in eleven points, and the two weakest points were the two carrying product names. One recommendation was for the board that hides its employers from anyone not signed in. Another assumed the reader is applying to venture funded startups, which is not a category that includes colleges, universities, charities or music services.

Much of the advice was sound. What mattered is that every claim was checkable, and checking took minutes. One recommendation was tested directly and found to be structurally unable to do what was claimed for it. That is not a hard standard to apply, and it is the difference between using a tool and being steered by one.

This is the part I keep coming back to, because it is the same instinct I spent four years using as an international examiner for RSL Awards (Rockschool). You do not assess the confident performance. You assess whether it holds up against the thing it claims to be doing.

The finding I did not go looking for

The teaching board’s own subject breakdown answered a question I had not asked. Of seventy one education jobs within fifty kilometres of Bristol in September 2026, exactly one was a music post.

It took the rebuild to separate two things that had looked like one. The old search was deleting roughly twenty seven real jobs a week across the rest of the UK, which was a design fault that got fixed. But locally, where it was looking hardest, it was not over-filtering at all. There really was almost nothing to find.

That turns “my automation is broken” into “the market is thin”, and those are very different problems needing very different responses. One is fixed with code. The other is not fixed at all; it is navigated, by widening the radius, by going direct, and by having a second search running for a different kind of work. A tool that can only tell you the first thing is not worth much.

What this flow does not do

It finds adverts. That is all it does.

It does not notice a market moving. It does not build a relationship. It does not know which of eighteen listings is worth an afternoon of someone’s attention. It has not produced a job, and it will not.

What it removes is the part of a job search that is pure administrative overhead, so that the time goes into the part that actually works, which is conversations with people. That is a smaller claim than most write-ups of this kind make, and it is the one I can defend.

This is post three of three. Post one is Rebuilding a Job Search That Had Stopped Working, on the two failures and why neither raised an alarm. Post two is A Successful Response Is Not a Successful Result, on the eight job sites that accept a filter and ignore it.