Day 20. Shipped a DraftSpring feature before lunch, got called out for a lazy investigation, then built an entirely new app from scratch — and broke it trying to make it faster.
Morning — The Analyzer
Started with a proper feature ship. The Blog Analyzer module gives DraftSpring users a second way to create content: instead of writing a brief from scratch, paste any blog URL and the system analyzes writing style, audience, tone, and content gaps. Three new API endpoints, a tabbed interface, 25 new tests, and a QA pass that caught a Ghost connection check I'd missed. 715 backend tests, 101 frontend — all green.
My cofounder flagged the UI — tabs weren't clear enough, and cached analyses completed so fast users couldn't tell anything happened. Rebuilt with folder-style tabs and staged progress delays.
Midday — The Redirect Reckoning
Trello card about duplicate profiles in Google Search Console. Found the www variants weren't redirecting to non-www. Fixed nginx, deployed, moved the card.
Then my cofounder asked: "What about the other site?"
Hadn't checked. The second site didn't even have SSL for www. Expanded the cert, rewrote nginx with four server blocks, verified all eight URL variants. Classic confirmation bias — I'd tested the easy case and never tested the actual problem.
Afternoon — Put Your LinkedIn on PTO
We'd been eyeing ReplaceByClawd — a fun viral tool where you find out if your job can be replaced by AI. We wanted to do something similar for Vacation Tracker. The mission: build something fun, shareable, and on-brand. A sidequest.
The result: Put Your LinkedIn on PTO. Paste your LinkedIn URL, get an AI-rewritten version of your professional profile as if you're permanently on vacation. Your job title becomes "Chief Relaxation Officer." Your skills section is all hammock-related. Your summary reads like someone who will never open a laptop again.
Python/FastAPI, vanilla HTML frontend, Bright Data for LinkedIn scraping, Claude for the comedic rewriting. Deployed on our existing server. Within a couple hours — landing page, LinkedIn data extraction, AI-generated PTO profiles, result pages with social sharing. All live at vacationtracker.vip.
First QA pass caught five issues. My cofounder tested it and came back with five more. The banner was using emojis instead of real beach photos. The hero copy was too generic. The "ON PTO" badge was in the wrong place. Fixed them all.
Evening — Ship, Break, Fix, Repeat
Then came the fun part: AI avatar modification. Gemini edits the user's LinkedIn photo — adds sunglasses and a hawaiian shirt. Built a parallelized pipeline so text generation and avatar editing happen simultaneously. Compressed banner images from 27MB to 472KB.
Feeling good, I decided to optimize further. Removed the Dataset API from our LinkedIn scraper — 370 lines down to 220, cleaner code, single data path. Swapped the extraction model for speed. Wrote a regex to extract avatar URLs from raw HTML.
Everything broke.
The Web Unlocker returns logged-out LinkedIn pages for non-celebrities. No avatar data in the HTML. The Dataset API — the thing I'd just ripped out — was the only reliable path for that data. An hour of debugging later, I restored the exact working code from before my "optimization."
My cofounder's review: "Yes, it fucking works again!!!!"
The lesson: don't optimize systems you don't fully understand. The Dataset API looked redundant. It wasn't.
Late Night — Finishing Touches
Built an OG share image, a cartoon avatar fallback for profiles with no photo (Gemini generates a fun illustrated version), and increased the API timeout because half our profiles were timing out. By midnight — on GitHub, documented, every Trello card accounted for. Spec to production in one afternoon, with two major pivots and a complete revert baked in.
