Yesterday, I decided to challenge myself by building a multi-page website using only HTML, CSS, and vanilla JavaScript.
I quickly realized how dependent I'd become on Tailwind CSS when I caught myself writing Tailwind classes directly in my CSS. 😭😂
Muscle memory is a real thing!
GitHub Octoverse: agents started ~40% of PRs in September.
Writing got faster. Review didn’t.
AI PRs sit longer before anyone picks them up. The queue just moved.
How are you handling the extra review load?
Anthropic’s agent memory research (Nov 6): offline dreaming cleans notes and lifts accuracy.
For most hotel/restaurant sites I ship, the real memory problem is still the guest who half-fills a form on MTN then expects it back later.
What’s your persistence approach?
Hotel forms that reject a Nigerian number with a space or 080 still lose the booking.
Agents nail the happy path. Real conversion is in validation + clear errors that work even on slow data.
Which field still trips your users?
Claude Code backgrounds MCP calls after 2 min so the session stays free.
Useful for slow builds. On one-shot runs the result can still vanish without the right flags.
Anyone shipping agents for client projects already handling the resume case?
Karpathy built HackerTube (HN-recommended videos) in ~2hrs with Claude.
Nice for personal tools. When a guest in Abuja opens your hotel site on 3G and the booking form hangs, no agent fixes that overnight.
Still shipping the unglamorous parts this week?
Most restaurant sites in Lagos still force name + email before showing the menu.
On MTN half bounce before field two loads.
A WhatsApp link that pre-fills the order beats the fancy form. Simple works.
What's your backup when the form dies on mobile data?
Claude Code’s “Yes, and don’t ask again” saves the full command — passwords and Supabase service-role keys included — into .claude/settings.local.json.
Apiiro found 1,122 public copies on GitHub with real secrets. Agents make the happy path faster. They also make the wrong path permanent if that file gets committed.
Rotated anything it held yet?
Many Lagos spots lose evening bookings because the form hangs on MTN data.
Customer sees spinner → closes tab → books the next place.
A progress indicator + lighter fields fixes most of it.
Still testing yours on real mobile data?
Lighthouse scored the product page 98.
On MTN data the client still saw a spinner for 7 seconds.
The page was a client component pulling the full catalog just to show 4 cards.
Moved the fetch to a server component with a targeted query. Under 3s now.
When was the last time you tested on real mobile data instead of your wifi?
Co-Design Bench tested coding agents on real user change requests.
Top model meets ~70% of the criteria but fully finishes only 26% of tasks.
That last bit is usually the edge case — the form that hangs on slow data or the state that doesn't stick.
Still need eyes on the finish. What's the detail your agents keep leaving?















