I Asked AI to Hack My Job Search. Here’s Why It Backfired
I’ve been plugging away at the job search in all the ways my career coach suggests for four weeks now: using tools to help tailor my resume to be ATS-proof, keeping on top of job notifications and applying promptly, doing the detective work to figure out who to follow up with and, most importantly, working my LinkedIn network. The applying to jobs online part is easy. But I ran up against an AI brick wall trying to execute my vision of setting up an AI agent to seamlessly and effortlessly identify open jobs I was uniquely qualified for before companies even posted them online.
What I tried
Knowing that I’d have more luck with a human who knew me in my corner, I asked Claude to comb through all of my first and second person connections and identify ones who worked for companies I’d enjoy working for. I thought Claude would know my likes and priorities well enough by now, but I added some details to help. Some conditions I gave included companies of 50-150 people with a focus on interesting science or technology (bonus points for something that could save the planet or improve society) who allowed jobs that were remote or were local enough that I could do hybrid. Truthfully, I had hopes that with all the information I’d already fed Claude about me, he’d just know what kind of company would appeal.
Turns out AI has some technical limitations (like privacy rules – I guess that’s a good thing). Also, no surprise), AI isn’t quite where I’d hoped as far as grocking me.
Damn those privacy rules
The gist of my prompt was “Can you scrape my LinkedIn at https://www.linkedin.com/in/j-allie-cliffe/ to identify connections who work for companies I might like to work with?”
Claude quickly responded “I can’t directly scrape your LinkedIn profile or access LinkedIn data through web scraping, as LinkedIn has strict policies against automated data collection and requires authentication to view most profile information.” D’uh.
Claude’s an overachiever, so of course he offered – unprompted – to “definitely help me with your job search in other effective ways”. He listed some tasks he could help with, ending with a list of “helpful” questions I could answer to help get started:
- “What type of role are you looking for, and in which industry or location?”
That’s the kicker, right? If I knew the job titles or industries, I’d be 90% of the way there already. I could list the obvious job titles, but what I was hoping for was some insight based on my previous work as well as… magic??
No worries. I carried on.
A focused try #2
I took up Claude’s offer to “Research companies in your field/location and identify if they’re hiring.”
I gave a longish prompt that included the size and type of company I wanted to work for along with a summary of some of my strongest and most transferable communications skills to use when searching. I added my location and the condition that it be either fully remote or hybrid + local enough that I could commute.
Claude gave me lots of results! The first few jobs I skimmed paid much, much less than what I need to make, however. Rookie mistake. So I added my salary requirements.
Claude cheerfully redid his previous effort, and came up with forty jobs and job titles that met the criteria, broken out by several categories (from ”Tech companies with confirmed high salaries” to “Mission-driven and climate tech companies”). Claude also included a few active job links from LinkedIn or Indeed, many of which I had not yet seen.
At first glance, Claude’s output seemed like it might be on target. Headers like “Mission-Driven & Climate Tech Companies” gave me hope. Exploring the output more carefully, I found he hadn’t exactly hit it out of the park. His recommendations for specific companies and exact job titles were bland and uninspired, the same “Googles” and “Nvidia”s that a green high school student might suggest. Delving into links revealed lots of flaws – either in Claude’s ability or my prompting.
- EGS Senior Technical Writer for a digital games store ($50/hour) was not the mission or salary I was going for.
- “Science writer/editor roles in environment/climate-change related subjects morphed into “Best remote technical program manager,” a mantle I’m *not interested* in taking up (and I’ve been clear about that in the training material I fed Claude).
- I’m not sure the list included a lot of companies that I know my connections work for.
Conclusions and next steps
Well that was disappointing. I keep reminding myself that Claude is still only a toddler, albeit one with a very big RAM – and I expected too much. By not understanding Claude’s limitations and not giving a specific enough prompt, I was setting Claude and myself up for AI failure.
I expected Claude to be more insightful, more introspective, more… human. And AI isn’t there yet.
I was hoping to avoid manually investigating every company that every one of my 600+ connections worked for myself. It seemed like the sort of task that AI does well, but I don’t believe that’s what Claude did.
Claude offer ed a long list of job titles, broken out in a way that made sense, but wasn’t super useful. Since it’s an LLM, they’re the most probable titles given the prompt and input data.
That’s worth something. I’ve had a hard time narrowing down my job title.
I learned I need a better way to keep Claude trained on my personal training data. Right now, if I start a new chat, Claude forgets everything I’ve fed him in a different chat. To keep from repeating the training, I’m doing almost everything in one mammoth chat and it’s super hard to go back and find things. I will need a better system for keeping track of Claude’s voluminous output, because my RAM space is limited!!
Also, I will need to think more carefully about how to get be more clear or more clever about my prompting.
All exercises for later in the week.