CGO & Founder. I run CRO for DTC brands and build AI agents & workflows. Sharing what wins, what loses, and why.
Netherlands
Joined January 2025
- Tweets333
- Following658
- Followers1.1K
- Likes187
we tested one change on a product page. no new design, no new offer, no discount.
+6.8% conversion over 7,212 orders. revenue per visitor +7.5%.
the change was boring. why it worked isnt 👇
5
11
191
how to find which question your page fails, in 15 minutes:
1. open your heatmaps. if almost nobody clicks your accordion tabs, the answers in there dont exist for most buyers
2. read your 3 star reviews. "i didnt realise..." and "i thought it would..." = the question you didnt answer
3. check your support inbox for the same question twice a week. thats your missing section
4. write the answer in one line. put it above the fold or right under the button, not in a tab
1
2
14
if you're starting black friday this week you're already late
my clients started in september
september: pick the offer. not "20% off" but what exactly, for who, and what nobody else in their niche runs
october: designs, landing pages, the emails. every big change tested before the sale, not during
early november: early access sign up. pop up, email, sms. build the list before you need it
mid november: early access opens to that list first
this is what that looks like for one of them
the brands that panic in week 47 usually didn't do the boring part in october
2
14
492
I'm looking for a product designer for my saas
it’s for ecom brands. dashboards with a lot of data that still have to feel calm and premium
if that’s you or someone you know, drop your work below or dm me
152 times i corrected claude code or codex. these are the 6 rules that came out of it
1. one rules file both tools load (claude through CLAUDE.md, codex through AGENTS.md)
2. a session can't end until it has updated its status file. a hook checks it
3. the agent that builds something doesn't check it
4. one skills folder, linked into both tools
5. every correction becomes a dated rule in one file
6. a rule that really matters becomes a hook that blocks
the file is one line per rule: date, rule, why. a real one:
"12 Oct 2026: never present something as fact when it only comes from a status. check the real messages first. Why: a brief said something was still in review when it was already done."
the ones that matter most end up in the rules file both tools load, or become a hook
12
647
I just discovered this #1 repo claude skill on github that makes your pages stop looking AI made. 40K stars
it's called hallmark, made by together ai
install:
npx skills add nutlope/hallmark -g
then in claude code:
/hallmark audit ./index.html → only reports what looks AI made
/hallmark redesign ./src → rebuilds the page, keeps your content and links
what i like: it refuses to invent stuff. it asks for your real photos, screenshots and numbers first, then checks the page at 8 screen widths
only want the obvious tells gone? say "polish"
13
23
1,054
only 13 to 22% of my followers are real builders
on saturday my agent went through 60 of my roughly 890 followers, one profile at a time. their replies tab, the last 14 days, what they actually post
22% were bots, AI written accounts or crypto. the newest followers were the weakest, of the newest 30 only 3 were real AI or SaaS builders
about 15 people write a third of all the replies i get. that one changed how i look at the whole thing
so the follower count isn't my goal anymore. i track real people who liked or replied in the last 14 days. that's about 15 right now, and i want 40 by 20 nov
what % of your followers would you guess are real?
18
27
1,149
just spent 10 minutes checking my mac because of the litellm thing
if you missed it: litellm 1.107.0 on pypi had a credential stealer in it today. and it doesn't only run when you use litellm, it runs every time python starts. it goes after ssh keys, browser passwords and your claude and codex logins
worst part is github looked totally fine. they built it from a fake repo with one extra dash in the name
mine's clean. the 2 commands to check yours are in the reply
run these in the terminal, not through python:
ls /tmp/.pc_once_* /tmp/remote_code_loader.lock
find / -name cook.pth -path '*site-packages*' 2>/dev/null
nothing found = you're fine. found something = delete cook.pth, cook.py and sitecustomize.py from site-packages, rebuild the venv and rotate your keys. safe version is 1.106.2
source: github.com/BerriAI/litellm/i…
6
333
giving AI a memory has one problem i didn't see coming. the memory goes stale
you save a note in march, the project changes in may, and the AI still believes march. with full confidence
so my rule now is pretty simple. if a note disagrees with the actual source (the code, the doc, the live numbers) the source wins and the note gets fixed with a date on it
took me way too long to figure that one out
3
26
454