my site scored 46 - so i gave Claude one prompt

from the episode: Anthropic is so far ahead it's not even funny - Claude now runs a team of agents | LAB 0022

Lenny's AI Builders - LAB 0022

four Claude agents, one prompt, 20 speed killers - and the worst one was embarrassing.

my own software's landing page scored 46 out of 100 on Google's speed test, on a phone.

so i gave four Claude agents one prompt, let them hunt down what was slowing it, and fixed it all in one pass. what they found was embarrassing.

why it matters

a slow page loses people before they read a word. Google's own research (with SOASTA, 2017) found that when a phone page takes 1 to 3 seconds to load, the chance someone leaves goes up 32%.

and you don't need to be a developer to fix it now. Anthropic just updated the Claude Code desktop app for running lots of agents at once - you can see which ones are working, which are done and which need you. that's what makes a job like this one prompt.

what you get

one prompt you can copy. it's Voxyz's "20 things that slow your site down", plus the line that makes Claude measure before and after.
four agents checking at once. images and fonts, frontend code, database and APIs, servers and network - one agent each.
real Google scores. Lighthouse, phone mode, before and after - not a vibe.

the real example: Threadify's landing page

i ran it on Threadify's landing page, in the Claude Code desktop app, on Opus 5.5.

the worst find: every person who opened the home page had to download the whole Threadify app - the stuff only paying customers use - plus a YouTube player and the payment form, before they clicked anything. it's like making everyone carry the stock room in before they can see the shop window.

the sneaky find: a celebration sound. its helper switched on the phone's whole audio engine the second the page opened, before anyone did anything.

on a copy of the site with every fix, measured side by side with an untouched copy (median of 5 runs each):

score: 48 → 86 out of 100
phone freeze (blocking time): about 4.2 seconds → 55 ms
page weight: 2.3 MB → 0.7 MB (69% lighter)

the live site still scores 46 until my teammate Sam ships the fix. 86 is the fixed copy, not the live site yet.

try it

update the Claude Code desktop app and open a new session in the Code tab.
pick your site's folder. work on a copy or a new branch, not your live site.
paste Voxyz's whole post (link) and then the prompt below.
let the four agents check first. they list what they find, file by file, before anything changes.
read the before and after numbers Claude gives you, then put the fixes live the way you normally would.

copy this prompt

paste this straight after Voxyz's post:

Lenny’s AI Builders / LAB 0022

send one subagent per group to check my site against the list above without editing anything. for each issue, note the file and the page it slows down. run the database subagent at high effort.

before changing anything, make a production build (or use a preview link), run Lighthouse 3 times in mobile mode on the home page and my 2 or 3 slowest pages, and note the median scores.

then fix everything from their lists yourself, measure again the same way, and give me the before and after numbers. don't remove features to gain speed. ask me before removing any dependency.

you'll know it worked when

you have a before score and an after score from the same test, run the same way - Lighthouse, mobile, 3 runs or more, median.

and every page still does what it did before. nothing got faster by being removed.

a bad result: one run each (scores jump around - mine swung from 28 to 49 on the same page), a desktop score instead of a phone score, or a feature that quietly stopped working.

limits and dates

one run, one site: Threadify's landing page, measured on 15 October 2026 with Lighthouse 12.8.2 in mobile mode.
a copy, not live: both sides were production builds on my Mac, run turn about. the untouched copy scored 48; the live site scores 46.
not shipped yet: the fixes are waiting for Sam to review and release. the live number moves when he ships.
scores swing: the before side ranged 28 to 49 across 5 runs while the Mac was busy. the after side stayed between 81 and 86.

LAB 0022 went public on YouTube on 16 October 2026 at 7:04 pm Perth time.

the rest of this LAB

Claude Code's desktop app runs your agents. Anthropic updated the app for running lots of sessions side by side, with a sidebar that splits them into needs input, working and completed. update the app and start a new session. my take: your laptop is an office now.
Grok Voice 3.0 tops the voice board. SpaceX SI says Grok Voice 3.0 scored 86.9 on Artificial Analysis's speech-to-speech index, ahead of GPT-Live-1's 83.2. the API is 8 cents a minute, the same as Grok Voice 2.0. GPT-Live-1 still wins on accuracy (74.5% vs 71.6%); Grok answers faster. my take: AI answers the phone now.
Grok Bot opens its connector store. founders can submit their product as a Grok Bot connector, an xAI engineer says Grok Bot now supports plus-aliases on its email, and Corey Ganim shared a prompt that plugs Claude and ChatGPT plans into Grok Bot (his prompt, not an xAI feature). my take: the plugin stores are the new app stores.
Grok Bot checked his accountant. investor Jason Freedman says he gave Grok Bot 4 years of tax files - 266 of them - and it flagged mistakes his CPA missed. that's his own account, and he invests in Magnetic Tax. my take: AI doesn't replace your accountant. it checks them.
AI cyberattacks within a year? Elicit's CEO Andreas Stuhlmüller says it's plausible we'll see AI cyberattacks on the power grid, banks, the internet or water within a year, and suggests a few weeks of cash, water and food. that's his opinion - nobody has proved the timeline. my take: back your shit up.
Claude Code stops torching idle sessions. Lydia Hallie at Anthropic says Claude Code now shrinks a chat to a summary after about an hour idle, so coming back doesn't re-read the whole thing and burn your usage. my take: finally, usage you can plan around.
Tibo's day 10: another Codex reset. Tibo says Codex usage got reset again on day 10 of OpenAI's 28 days of ships - after day 6 and day 8.

sources

next step

watch LAB 0022 on YouTube - the practical starts at 11:31.

ran it on your own site? drop your before score under the video.

want the next one? get the next field notes. optional - these notes stay free to read.