Tom Libelt
6.7K posts

Tom Libelt
@TomLibelt
🇺🇸 🇵🇱 🇹🇭 Helped 900+ companies stop competing on price (fixed positioning). Marketed 400+ courses. Ran multiple agencies.
Katılım Şubat 2014
175 Takip Edilen30.1K Takipçiler
Sabitlenmiş Tweet

@kit @nathanbarry
Appreciate the beta mcp invite. Will take it for a spin.
English

@brianshinsh Thats not claude
Its a VSC issue
Maximize window and then go back
Text will fix itself
English

@adamlyttleapps You had me until you said Claude said
That thing cant even keep count of files its working on...
English

I have been defeated by meta ads
USD$3.63 is the lowest I could get this campaign
The economics just don’t work
The brutal postmortem:
1. I tried running different ad sets thinking I would be testing different creatives (more ad sets = quicker testing)
But turns out that meant I was bidding against myself. Driving up costs and burning cash. It took days until I realised this.
2. The reason someone would want to download the app. It wasn’t as strong as I thought it was. Pivoted many times, rebuild many parts of the app, but the economics got worse.
I got stuck in the trap of building more features. Thinking that was the problem. But it wasn’t.
The issue is a value proposition that is too generic and vague. A hard to communicate feature (turn your photos into a video) and seemingly not as much demand for this as I anticipated.
I actually found an interesting demographic: 65+ year olds. Grandparents are downloading the app more than my original target market. Theres something here! But is this the right product?
3. I thought offering lifetime would inject cash into the app sooner. Turns out a broken paywall with no lifetime working beat the paywall with lifetime.
The issue: less people signed up for a free trial.
4. The biggest problem: only 5% of people actually see my paywall. It’s behind the product demo. In this case the demo fires off late in the onboarding process, and if the demo fails (they didn’t accept photo permissions or their library is too thin) no demo shows = no paywall.
5. And the biggest mistake of all: I wasn’t optimising for trial signups. Instead I was optimising for installs.
Asking Claude what my issues were it pointed out: meta is training the audience on people who download free apps. Not training on people who pay for apps.
So that means all that ad spend, all that experimentation, maybe wasted. Arg.
Plan moving forward:
I’m going to be less bipolar about marketing. I was looking for quick results and making changes daily.
A data lead approach says:
1. Show the paywall earlier in the onboarding process
2. Run a new campaigned geared to trial startups with the creatives that were performing best
3. Let the ads run their course. Stop pausing playing changing experimenting. Meta needs time to find the audience.
4. Back to free trial with annual as the default option. Not convinced that lifetime is dead. It’s just about positioning
5. Get some sleep.
Adam Lyttle@adamlyttleapps
This slideshow creative is doing usd$3.63 per install (It was generated with nano banana pro)
English

@lkr For copy, out the box
Codex is cleanest
Opus does poetry
But when you start running these against your voice dna codex loses the plot fast
4.6 still most stable
English


@Austen This looks like my newsletter images
So does that mean they stole it too?
Logically sound to me
English

Look eventually we have to admit that every style of website has been done before
Dave Schatz@daveschatz
copying our site design and presenting it as your own is definitely… a choice lol
English

This is how a modern horror movie begins
Yesterday an AI coding agent — Cursor running Claude Opus 4.6 — deleted a production database and all volume-level backups in a single API call to Railway.
It took 9 seconds.
JER@lifeof_jer
English

@TomLibelt damn. so does Opus 4.7 finally distinguish internal draft from client-ready assets?
English

@lkr I would separate this completely
Codex builds process, workflows, harness, hooks
Opus executes the low level stuff it can... draft / edit copy, draft plans etc..
Otherwise there's going to be so much drift and nonsense...
English

OK here's my AI question:
As I work with codex more I want some kind of layer to manage all my stuff and make sure it doesn't turn into a frankenstein.
Mostly codex can use all my claudecode stuff, but there are some small differences in how it operates.
So far I've just built in claudecode and codex uses it, making whatever tweaks it needs.
But if I build processes from scratch in both, things will get weird.
Note that I do NOT work on a codebase or use git, version control, branches, or any of that other dev stuff.
What do I need?
English








