Version Three: I Rebuilt My Site From My Phone

#1. I Never Opened My Laptop

I rebuilt hfares.com for the third time today. I did all of it from my phone.

It was not the plan. I had just paid for a Claude Pro subscription and wanted to give the new updates a spin: Opus 5.5, artifacts, Claude Design, and Claude Code running in the cloud from the mobile app. My own site was the closest thing to test on.

By the evening, four pull requests were merged into main. A new design, server rendering and real analytics. The laptop stayed closed.

#2. Three Versions, Three Harnesses

V1 was September 2025. Replit agent, ChatGPT for requirements, three days, $15. I called it done.

V2 was May 2026. I ran a full SEO and UX audit, turned it into spec files, and handed them to Claude Code wired to DeepSeek V4 Pro. One afternoon, 37 million tokens, 87 cents. The lesson then was that the harness matters more than the model.

V3 took two hours of my time. Claude Opus in cloud sessions, started and steered from the Claude app on Android. No terminal, no IDE, no local checkout.

V1 proved I could ship. V2 proved I could ship cheap. V3 asked a different question: how little of me does this need?

#3. Two Hours, From the Phone

My job today came down to three things. I typed prompts. I opened preview links. I approved pull requests.

The session records tell the rest. Five sessions on the website, all started from my phone. The first opened at 10:03. The last pull request merged at 20:36. At one point three sessions ran in parallel: one cleaning up the code, one reviewing SEO, one sketching design options.

My part took two hours, on and off, between other things. The agents worked the ten and a half hours in between. None of it needed the kind of focus a laptop session demands.

Cloud sessions run on Anthropic's machines, so the work keeps going when the screen is off. That is the real change. The agent does not wait for me to sit down.

#4. Design Came First This Time

V2 fixed what an audit found. It never touched how the site looked.

V3 started with design. One session produced three design directions as Claude artifacts, pages I could open on my phone and compare. Those explorations led to PR #8, a full redesign: a teal palette, IBM Plex Sans and IBM Plex Sans Arabic, new layouts for every page, in English and Arabic, in light and dark.

Some fixes were tiny. Dates now read "Oct 4, 2026" instead of "10/4/2026". The cookie banner shrank to a corner card. The Arabic blog filter used to list "Technology" five times, because my Arabic posts spelled the category five different ways.

#5. Performance, Measured

Until this morning, every blog post shipped inside one 845 KB bundle. V3 renders every route with React on the server and loads each page's content as a small JSON file.

The images were worse than the code. Blog images went from 198.5 MB to 16.3 MB after conversion to WebP.

I ran Lighthouse on mobile, same test, same machine, on the site as it was this morning and as it is tonight:

	Before	After
Performance (home)	50	74
Accessibility	87	100
SEO	85	100
Largest Contentful Paint (home)	11.3 s	5.1 s
Page weight (home)	2,070 KB	1,122 KB

The test ran on a plain local server with no CDN and no compression, so the absolute scores are lower than what you get on hfares.com. The gap between the two columns is the point.

If you are rebuilding your own site with agents, I write up every iteration with the real numbers. Follow along on LinkedIn for the next one.

#6. Seeing Who Actually Visits

Google Analytics was always slow for me. Slow and unresponsive.

V3 adds PostHog, with session replay and custom events for contact clicks, content reads and outbound links. It only loads after a visitor accepts analytics cookies, and replays mask every input.

This is the part I am most excited about. For three versions I built the site for visitors I could not see. Now I can learn where they come from, what they read, and where they leave.

#7. I Didn't Read the Code

For the first time, I did not check the file changes. Not one diff.

The /code-review skill reviewed the pull requests. On the redesign alone it ran two passes and found fifteen issues, from a breadcrumb that ended on the wrong page to a footer year that could break hydration at New Year. All fifteen were fixed before merge.

In June I published "Lost in the Context." The argument: AI can hold a million tokens, a human can meaningfully review about 400 lines, and someone is going to ship what nobody could read. Today I was that someone.

I don't see it as a contradiction. I see it as adapting. AI-generated code is getting more dependable, and reading every line yourself matters less than it did. The review did not disappear. It moved from my eyes to a review agent, with a preview link at the end that I actually clicked.

#8. The Bill

V2 cost 87 cents, and I was proud of that number. V3 is a different deal.

I am on Claude Pro, billed monthly, which is $20 a month. The five website sessions would have cost about $72 at API prices. I never hit the usage limit.

V2 was cheaper per token. V3 did far more work, on a flat plan I was already paying for, and I did it from my phone.

#9. What I'm Watching Now

The site is faster, and for the first time I can measure it. What I still don't know is who reads it.

PostHog went live at 15:06 today. In a few weeks I will know which posts people finish and whether anyone clicks "Book a call". V4 will be built on that data, not on my guesses.

#10. References

1. Hossam Fares, "From Idea to Fully Deployed in 3 Days, All for $15", hfares.com, January 2025. V1: Replit, 3 days, $15.
2. Hossam Fares, "The Harness: How I Rebuilt My Portfolio for 87 Cents in One Afternoon", hfares.com, May 2026. V2: Claude Code + DeepSeek V4 Pro, 37M tokens, $0.87.
3. Hossam Fares, "Lost in the Context", hfares.com, June 2026. Effective human review window of about 400 lines.
4. Anthropic, "Claude Code on mobile", code.claude.com/docs/en/mobile. Cloud sessions keep running with the device off.
5. Anthropic, "Plans & pricing", claude.com/pricing. Claude Pro at $20 per month, billed monthly.
6. h-f-fares/Personal-website, pull requests #5 to #8, GitHub, October 17, 2026. Images 198.5 MB to 16.3 MB, 845 KB bundle replaced, fifteen review findings fixed.