Tim Grassin

3x exited founder in Bangkok. Building @roxskinapp in public: 3D avatars that help amateur Hyrox athletes get sponsored 🏁

[Tim Grassin @timgrassin]

3 design calls Claude Code couldn't make for me on my marketplace:

1. Shiny skin on the 3D avatar. My verdict: "looks like a doll." Gone.
2. Black and grey theme: "too generic and boring."
3. Multi-colour theme: "too flashy."

We landed on white, blues for data, orange only for "needs work."

AI gives you options in minutes. Picking is still your job.

[swyx @swyx]

“if you want good taste, you have to eat”

- @jxnlco

[Tim Grassin @timgrassin]

3 things I deleted from my AI-built marketplace in its first 3 weeks:

1. The approval step on new auctions. One athlete waited 36 hours for me to click approve.
2. 2 homepage sections I'd just built. They didn't earn their place.
3. The sponsor pitch on the homepage. A friend said pitching athletes and brands at once was confusing. It's athletes-first now.

With AI, building is cheap. Deleting is the fastest thing you can ship.

[Tim Grassin @timgrassin]

The context Claude was missing on my build: how big I wanted it.

I asked Claude Code for a directory of HYROX gyms. It came back with 61.

Me: "How come you limited yourself to that many?"

It had been careful. Small batches, one gym at a time.

Once it knew I wanted every gym in every race city:

1. Scrape Google Maps
2. Agents check each gym's own website
3. No quote from the gym's site, no listing

861 gyms in 79 cities. About $5.50.

[Alex Albert @alexalbert__]

A lot of getting good work out of Claude now is just finding the context it’s missing and figuring out how to close that gap

Claude Meetings is one way to bring in the remaining slices and has proved very useful internally

[Tim Grassin @timgrassin]

My social media runs on 2 AI agents and 1 Google Drive folder:

1. Claude reads my 82 guides, makes the videos and writes a week of posts
2. It drops them in a Drive folder
3. A Grok agent picks them up and schedules them in Buffer
4. I approve. That's my whole job.

Instagram, X and YouTube. I write none of it.

The trick is a one-page rules doc both agents follow.

[Tim Grassin @timgrassin]

Agree. I built a marketplace with Claude Code in 21 days and 324 prompts.

The prompts were easy. The attention was the work:

1. Noticing the first 3D avatar looked like a shop mannequin
2. Spotting an athlete's auction stuck "In review" for 36 hours
3. Watching 2 friends get stuck at sign-up, then turning every bug they hit into a permanent test

AI cut the cost of building. It didn't cut the attention.

[Guillermo Rauch @rauchg]

PSA: creating wonderful products still requires extraordinary amounts of human focus, effort, motivation, and love. And an attention span.

[Tim Grassin @timgrassin]

Marc Lou got $112,062 from brands for one race. He has a big audience.

An amateur athlete doesn't. So who sponsors them?

1. Their gym
2. Their physio
3. The café that sees them every morning

Brands don't browse websites looking for amateurs to sponsor. They back people they already know.

That's why spots on RoxSkin start at $50.

[Tim Grassin @timgrassin]

The boring half of building RoxSkin with AI:

1. 98 race pages, each checked against the official race page
2. 82 guides
3. Every real bug turned into a permanent test
4. One avatar prompt rewritten until the skin looked real

The fun half: 7 visitors on the first day of analytics, 145 on the best day.

[@levelsio @levelsio]

"Boring & fun" is a great way to describe building with AI

[Tim Grassin @timgrassin]

Visits to https://t.co/svUqJ9BK9F from ChatGPT:

Last week of October: 43
First week of November: 79

On a site I started on 14 Oct.

If you only track Google, you're missing a channel that's already sending people.

[Tim Grassin @timgrassin]

@carloAI Pledged on the spot. Hope it actually makes it to me by Feb 🤞🏼

[Tim Grassin @timgrassin]

After my homepage, the page Google sends the most clicks to is a guide on how to volunteer at a HYROX race.

10 clicks in 5 days, from 331 impressions.

Another guide, racing HYROX for charity, pulled 494 impressions.

Your customers search for the jobs around your product. Write those pages.

[Tim Grassin @timgrassin]

@bird_helen67403 They landed on a page in my site after mention on gpt

[Tim Grassin @timgrassin]

My 3-week-old site was already getting visitors from ChatGPT.

RoxSkin in its first 21 days: 4 visitors from ChatGPT, 4 from Gemini. The first one came on day 4.

What I set up for it:
1. llms.txt
2. Clean structured data
3. Answer-first opening paragraphs
4. Bing plus IndexNow after every deploy, because ChatGPT search leans on Bing

Small numbers. Claim your spot before it gets crowded.

[GREG ISENBERG @gregisenberg]

openai just let every business claim a profile inside chatgpt. my honest reactions:

1/ this could be the google my business moment for ai, and a lot of agencies made a great living (millions!) in the 2010s just claiming and fixing listings for local businesses

2/ ads can now click through to a conversation with your agent instead of your website (interesting)

3/ every business agent runs on context the company writes, and most brands are going to be bad at it

4/ your product feed is your new seo. boring, and it matters more than your blog now in some ways if you think about it

5/ the dashboard tells you how often chatgpt mentions you but HIDES the prompts and your competitors, so you still can’t see why you’re winning or losing

6/ discovery topics shows what chatgpt already cites you for, which is basically a free content calendar

7/ agents are only open to us ecommerce and online services right now. plumbers, dentists etc are next, and so lots of opportunity to help main street here

8/ businesses can’t see the conversations, so nobody knows which chats turn into sales yet

9/ i’m not sure people want to talk to a brand’s agent. i think they’ll want their own agent to talk to it for them

10/ 1.2 billion people use chatgpt every week. i’d go claim my profile before finishing this post if you can

[Tim Grassin @timgrassin]

I already skip code review. I can't read code.

What I review instead on RoxSkin:
1. Session recordings. People were tapping things that didn't respond
2. Two friends signing up. Both got stuck in onboarding
3. Any step where a user waits on me. One auction sat "In review" for 36 hours, so I deleted the approval step

Code is Claude's job. Behavior is mine.

[Tony Dinh @tdinh_me]

Very cool, but maybe in about 12 months from now, we won't do code review anymore.

Things left for human to review is product behavior / UI (taste)

[Tim Grassin @timgrassin]

How to copy someone's success, in 3 steps:

1. Copy what works. I copied Marc Lou's 3D avatar, the doubling auction and the race-day tattoos.

2. Find the one part you can't copy. For me it was Marc's audience.

3. Solve that part. It's the part that decides if your copy works.

[Tim Grassin @timgrassin]

@dada_6 Lol this is the most random room ever

[Tim Grassin @timgrassin]

Google clicks to https://t.co/svUqJ9BK9F:

Last week of October: 86
First 5 days of November: 132

Best day so far: 2 Nov. 38 clicks from 3,047 impressions.

Every page came from a prompt. Free race pages, gym pages and guides.

[Tim Grassin @timgrassin]

Saw this on a small scale building RoxSkin with Claude Code.

The guides were written in parallel. One agent per race, about 8 per turn.

What broke:
1. Web search caps at about 200 calls per turn
2. My QA server and preview shared a folder and corrupted each other

Both were my setup, not the model.

[Guillermo Rauch @rauchg]

At @vercel we got to test this model before release. Best way to summarize it: astonishingly brilliant.

One interesting takeaway: it's very clear that this new generation of models can take on unprecedented bodies of parallel work.

To actually seize this intelligence, you need cloud sandboxes. Full stop. The model can delegate and orchestrate far more than a single computer can handle. Likely far more than you can even imagine it can handle.

Your bottleneck is now ① your ambition and ② your parallel computing power.

Btw, I learned from @cramforce you don't need to overcomplicate it. Install Vercel CLI, setup the Plugin and tell the agent to use /𝚘𝚏𝚏𝚕𝚘𝚊𝚍-𝚝𝚘-𝚜𝚊𝚗𝚍𝚋𝚘𝚡 skill. It'll provision and de-allocate cloud sandboxes as needed. You don't need a "multi-agent remote harness ralph orchestrator" 😆

[Tim Grassin @timgrassin]

Mine is maintained by a founder who can't read code.

RoxSkin: 324 prompts to Claude Code in 21 days.

What keeps it alive:
1. Every bug a real user hits becomes a test that runs forever
2. Tests run on a copy of the database, never the real one
3. I only approve what's irreversible

[Amjad Masad @amasad]

"But who maintains vibe coded apps?"

How about self-healing software...

[Tim Grassin @timgrassin]

"Can we explain to athletes that they should not expect brands to find them"

The most important prompt of the month, and it came 20 days in. More on that soon.

[Tim Grassin @timgrassin]

I read back all 324 prompts I sent Claude Code while building RoxSkin in 21 days.

The prompts that mattered most pushed back on the output. A thread of the best:

[Tim Grassin @timgrassin]

"I've submitted a few SEO related requests, giant prompts. Could you turn those into a skill so that I can apply to other projects?"

Turn every good playbook into something reusable.