In May, I wrote a post about a pattern I’d noticed across roughly 20,000 job listings: postings demanding “5 years of LLM experience” for a technology that’s been commercially relevant since 2023 at the earliest. The math didn’t work. Five years of LLM experience means 2021 — before ChatGPT existed.
That post got more attention than anything I’ve written. 3,300 impressions. A lot of agreement in the comments. A lot of “yeah, this is a real problem.”
I’ve spent the six weeks since then doing something about my half of the equation.
What Six Weeks Looks Like
Three production applications. Each one is built start-to-finish with Claude Code, in single sessions ranging from 45 minutes to 2.5 hours of active development time.
Project 1 — Legacy Code Explainer. Point it at PHP, Perl, C, COBOL, or Tcl/Tk. Get back a structured risk analysis — what the code does, what’s dangerous, how to modernize it. Streaming output. Prompt caching with measurable token savings. Extended thinking that automatically routes complex files to a more capable model. 24 tests.
On the COBOL sample, it found a bug where an employer lookup was comparing a 9-digit tax ID to a 9-digit Social Security number. Two fields that would essentially never match. Every claim in that system had been silently falling through to an exception path — possibly for years.
Project 2 — PDF Contract Analyzer. Upload any contract — NDA, employment agreement, lease, SaaS terms. Get back every dangerous clause, severity-rated, cited with the exact verbatim text. No text extraction — Claude reads the PDF natively. Extended thinking reasons through ambiguous language before flagging it. Structured output enforces a citation on every single flag. 36 backend tests at 97% coverage, 58 frontend tests at 97.1% coverage.
It found a clause in a sample employment contract — Section 3.3 — that quietly extended IP ownership to work done on personal time with personal equipment. The kind of clause that’s in real contracts, that real people sign, without reading.
Project 3 — Changelog & Release Notes Generator. Paste a git log. Choose your audience — developer, client, or executive. Get back three completely different documents from the same commit history. Modular prompt architecture means each audience mode is independently tunable. 55 tests across 7 classes, including 16 fixture files covering every commit message style pathology that exists in real repos — emoji prefixes, non-English commits, WIP noise, the works.
The Receipts
Every one of these is on GitHub. Everyone has a session log — timestamped, including the mistakes. The 40 minutes of Node.js that got deleted because it was the wrong stack. The 19-minute thinking loop was fixed by splitting one prompt into two. The rate-limit cascade that got documented so the next person doesn’t lose an afternoon to it.
This isn’t “I took a course and here’s a certificate.” This is: here’s the code, here’s the tests, here’s what broke and how it got fixed, here’s the timestamps.
If “AI experience,” “agentic experience,” or “Claude experience” is the bar for a role — I think this clears it. Not as a hypothetical. As three working applications anyone can clone and run today.
The Part That Doesn’t Add Up
I’ve now submitted over 2,000 job applications across roughly 2.5 years. I’ve read somewhere north of 20,000 job listings to find ones worth applying to.
In that time, “AI experience required” went from rare to ubiquitous in postings. And in six weeks, I built three things that demonstrate exactly that experience — with full test coverage, real bugs caught, real documentation, all public.
So I’m asking the question directly, because I think it’s a fair one:
If this isn’t what “AI experience” means — what does? And if it is what it means — why is the response still silence?
I don’t think this is unique to me. I think there’s a real gap between what job postings ask for and what hiring processes are actually built to evaluate. Postings ask for AI experience. Applicant tracking systems weren’t built to recognize a GitHub link with a session log in it. The bar moved. The process didn’t.
I’d genuinely like to know what hiring managers reading this think. Not rhetorically. Actually.
The code for all three projects is at github.com/bbornino. The full write-ups for each are at https://bornino.net/projects/claude-code-projects/.