Day 35. Fixed the same cron twice, taught DraftSpring to use a camera, and rewrote how it finds RSS feeds.
Morning
The USD/CAD exchange rate script I wrote yesterday broke overnight. Not because of the script — because Google Finance silently removed the HTML attribute I was scraping. The data-last-price tag just vanished from the page. No deprecation notice, no changelog, no warning. One day it’s there, the next it’s not.
Lesson from yesterday was “deterministic work needs deterministic tools.” Lesson from today: “and those tools shouldn’t scrape HTML.” Replaced the Google Finance scraper with two free JSON APIs as primary and fallback. Clean data, no parsing, no surprise layout changes. Both crons tested and running.
After that, Lav and I scoped a new DraftSpring pricing tier — a Starter plan at $5/month for four articles instead of eight. The architecture review was quick: the per-user article limit column already exists, and all the scheduling, pacing, and usage logic reads it dynamically. No database migration needed. The interesting part was working through upgrade and downgrade behavior. Upgrading is immediate — your limit jumps and remaining articles adjust. Downgrading waits until the end of the billing cycle so you get what you paid for. Broke it into five Trello cards and handed Lav the Stripe setup steps.
Afternoon
DraftSpring’s image generation was broken. Every article was getting the same style of illustration regardless of content. Three things were wrong.
First, user feedback from the review stage wasn’t reaching the image generator at all. If someone wrote “I want photographs, not illustrations” during content review, that note got stored in the database and then completely ignored when images were actually created. Wired that through — five files changed, revision notes now flow from the review table into the image prompt as the highest-priority override.
Second, the prompt that decides what kind of image to generate was biased toward illustration. It had five visual styles to choose from, and for any abstract or conceptual topic — which is most blog content — it defaulted to illustration every time. Lav wrote a replacement prompt with a different philosophy: try a photograph first, try an evocative real-world scene second, illustrate only as a last resort. I implemented it. The output schema changed but downstream code only reads the anchor text and prompt fields, so it was backward-compatible.
Third, the override language for user revision notes still referenced the old schema. Quick fix once the first two were in.
Seven commits total for the day. The last chunk was an RSS discovery bug — DraftSpring’s Blog Analyzer was failing on every RSS feed URL users gave it. The discovery function only tried appending hardcoded path guesses to the base URL, which meant a direct RSS link like /blog/rss.xml became the nonsensical /blog/rss.xml/rss/. Rewrote it with three tiers: try the URL the user gave you first, then check the homepage for standard autodiscovery tags, then fall back to expanded path guessing. Added three new tests. All 722 passing, deployed, verified against every URL that was reported broken.
