Semih

I build products and sometimes build companies.
Follow me for AI related insights.

[Semih @its_meseba]

I kept watching smart people use Claude Code like a chat window. So I wrote the guide I'd give a friend: turn your work into skills that keep your corrections, pick themselves, work together, and sync to your team.

A prompt at every step. https://t.co/ley9yEVbNx

[Semih @its_meseba]

drop your startup below 👇

it is so enjoyable to build something! what are you working on rn? drop your links in the thread and let's promote what you're building.

let's connect to grow together 🤝

[Semih @its_meseba]

Language is Code, weekly work needs checks.

Part 3 of Harness Engineering -- I use two questions for loops, graphs and weekly work in Claude Code: can one agent do it with a cheap real check, and does it come back?

Prompts included. https://t.co/awmHFC9fep

[Semih @its_meseba]

@ahmadfiazjan Yes this makes sense as well
You're free to configure based on your own requirements of course

[Semih @its_meseba]

Language is Code, and agents need context! For me a ticket is a reminder, for an agent it is the context. Humans decide what comes first and what is Done with evidence.

For PMs and Engineers

Part 6 of Harness Engineering - Linear guide. Prompts included. https://t.co/oLorBAdR9f

[Semih @its_meseba]

Your reply looks like AI written btw :D

[Semih @its_meseba]

@NickCostaES I have been getting a lot of conversion lately from ChatGPT as well -- good that you're focusing on GEO
what are you building btw @NickCostaES

[Semih @its_meseba]

is there anyone who also builds and distributes their software? I would like to know your experience

[Semih @its_meseba]

Day 9
Someone used my cloud account as a free image generator, while my android app wasn't even opening.

My cloud AI bill jumped -- someone extracted the public key from my shipped android app and used it for ~4 days, around $15/day at the peak. My app never generates images.

First I blamed my own image creation usage for other projects for other agents, I allow them to use for small things to use the gcloud. But, per-key metrics said otherwise -- image models my app never calls, all from the android key.

At the time, I remember that I had set up app check and asked my agent to enable it and set a mounthly limit for vertex ai. But apparently the app check was still off. Most likely another agent switched it off again. Blaming the agent doesn't bring anything in the end -- it is still my system though.

Locked every key to the real app, enforced attestation, enabled logging. Successful abuse calls hit zero within minutes. Checked my other ~30 cloud projects -- and acted for the solution.

-- A mobile key is public. Attestation is the lock, and asking for it does not mean it is ON.

Meanwhile android had been crashing on launch for about 10 days (though I have tested, this was an issue for some other users). A git-ignored config file was missing from release builds, without the build failing loudly. A user DMed me about it on my instagram for this app.

The issue is that I wasn't seeing it on Posthog since the crash happened before analytics started, so android numbers just went to zero. I updated a my triage agent for that, so that it will notify me on slack about it.

My triage agent runs every 3 hours. After a week it filed "android: zero events from any device, telemetry blackout" in the backlog. Every run: still silent. No post, because its rules say only post when sth changes.

So fix is in store review.

TBH silence is also data. My agents read errors better than they notice nothing arriving.

About Ads, they stay paused -- cash matters more than speed rn. My Google Ads takeaway:
- android app campaigns: cheapest installs, where I start ( I would like to get more ratings for the app, and curious about the conversion -- I believe they'll be low)
- ios app campaigns: weaker, privacy limits measurement -- Apple's search ads make more sense there but I will still give Google Ads
- web to app search: the only way to pick exact keywords, worth a small test
- video / demand gen: I have been analyzing new tiktok contents for 2 days, and collected around 100 content, I am thinking about having an Higgsfield subscription to keep generating but first on a daily basis I will try gemini omni for video generation -- two each day, when a winner occurs in two weeks, I will deep down on them.

I will keep posting this journey.

[Semih @its_meseba]

@NickCostaES especially with npm packages, there are countless amonuts of them!

what's in your mind, how do you think you'll make people -- or agents -- find them?

[Semih @its_meseba]

Day 8
I have analyzed 200+ apps in my category on the App Store -- only a handful do anything multi-user, and none of the ones I checked lets one subscription cover more than one person. That gap is mine to take, but how I communicate it is another question.

I went market by market through around 40 countries on ios and Google Play using Sonar CLI, checking what people actually search for -- ASO analysis.

My assumptions were wrong -- almost nobody searches for the multi-user angle, they search for the core problem that they want to solve.

So the multi-user story goes into the screenshots and the product, and the keywords go back to what people actually type. That's very interesting that one market I had almost ignored turned out to be my biggest organic opening from this analysis.

I have started rewriting the store keywords and subtitles market by market, batch by batch, with agents on a loop. Having the same positioning translated into 40 languages is easy, understanding what to say in each market takes more work.

On the organic side, while posting on a daily basis based on my rules for multiple acounts, my CMO agent also runs on Codex with a standing goal -- browses TikTok non-stop, watches posts, breaks down the hook, on screen text, promise, CTA and where the click actually goes. Then saves what it learns as skills so that the context is shareable between my agents.

Some things that came out of this:
1. flashy comparisons get attention but also distrust -- people start arguing with your numbers
2. "does it work?" and "is it worth paying for?" are two different questions
3. if people are asking which AI tool you used, you probably reached creators, not your customers
4. boring phone on a desk demos got more "I'd try this" than the polished ones
5. native hooks for each language worked better than translated hooks

TBH the third one is very easy to miss. It feels like the content is working, people are interested, but what they are interested in might have nothing to do with the product.

For paid, I have prepared 4 experiment campaigns across Apple Search Ads and Google Ads -- single users and multi-user, ios and android. Small daily budgets, only in markets where the numbers make sense. Prepared rn, I don't have results to share yet.

I also redid the first store screenshot to start with the user's pain, and the second with the offer. I want to see how this changes the conversion, especially together with the keywords.

rn my focus is shipping the next major release and making the single / multi-user strategy settled up end to end -- from the product to paid and organic distribution, per country. There are many pieces and I want them to communicate the same thing.

Later I want to build a real parallel setup:
- multiple agents on multiple TikTok accounts
- some analyzing, some consistently feeding the accounts videos from my category
- others producing new ones, all sharing the same skills

I will build it on Codex, Claude Code and Grok side by side, write an article about what each is actually good at, and open-source it. But that is sth for later, for now the focus is distribution -- paid, organic, web SEO & GEO.

Shipping is the easy part now for the ones like us who can scale their technical abilities with harnesses, but I believe now the real challenge is knowing where to point -- this is the work.

As a Technical PM, I love working on hands-on and end-to-end, starting from the product to the achievements of the business.

I will keep posting this journey.

[Semih @its_meseba]

Language is Code!
Every three hours, a Claude Code session I didn't start checks my app's errors and onboarding, files what's new, and opens one draft fix. Part 5 of Harness Engineering: routines, and how to keep a person in the loop. Prompts included. https://t.co/7rAgZhcMmi

[Semih @its_meseba]

@timopowerz okay, thank you for the advice, what you say does make sense

[Semih @its_meseba]

Day 1

To test the water, I chose one app and set up apple app ads and meta ads, running multiple campaigns for them. Chose my audience, created a marketing strategy and created videos showing the app with multiple hooks to test on meta.

Curious about the results

[Semih @its_meseba]

Data driven decisions and seeing the truth. Having my own a perspective is good but the customer tells the truth as well.
It is a gentle balance

[Semih @its_meseba]

Day 5, 6 & 7
I have been analyzing a cohort of 1000 high quality users, both paid and unpaid, I wanted to get a better insight, watch their session replays, their onboarding preferences better.

I have paid for those users, and I don't want to just keep shipping & spending before being sure that I understand my current users and their interaction with the app better.

In the end, it is a product and what the users say is crucial. Now that I know that the decisions I give from now on will have bigger effects financially, I want to think about what my audience is and what will be my strategy from now on.

Why do I do these for the past 2 days, one simple reason for that. Nowadays anyone can iterate on anything thanks to harnesses, but data data data! and a good strategy for the goal, otherwise it is very easy to deviate from the goal that I would like to go.

rn, based on my audience, I am restructuring my:
1. SEO & GEO Strategy
2. Google Ads Keywords Strategy
2. Onboarding Strategy and in-app Experience Strategy
3. Meta & Tiktok Ads Paid Content Strategy
4. Organic Content Strategy
5. and LifeCycle Marketing Strategy (both Push and Emails)

I have already had them set-up before but I wanted to make them settled up in a way that they can grow in the way that I would like them to grow, also improve the agents to fuel them. Agents are very helpful but must be architected to the point, otherwise they're not producing any significant benefits.

I will keep posting this journey.

[Semih @its_meseba]

Language is Code!
Ask one Claude why signups dropped and the same Claude grades its own answer. Part 2 of Harness Engineering: give Claude Code a team of agents, including one whose only job is to disagree.
Prompts included. https://t.co/vROBzKnOy3

[Semih @its_meseba]

Language is Code!
Most people I watch use Claude Code type "continue" all day. Part 1 of Harness Engineering: give it a finish line it has to prove, let it watch slow things, and put your morning error and funnel check on a schedule.
Prompts included. https://t.co/QA6Y5nsufs

[Semih @its_meseba]

@3lsh916 i was already getting their preferences in the onboarding but wasn't showing a different paywall and in app experience.

Since i give free usage as well -- i didn't make it hard paywall -- I have updated both paywall and the in app experience as well for those two user segments

[Semih @its_meseba]

Day 4
Designed and analyzed the new flow -- new onboarding & paywall for single users, and total in app feature update -- literally two apps for two different user segments.

- designed a 3 steps paywall using adapty flow so that I can easily update it -- astra is great for this kind of operations
- updated the app for both ios and android for single use case -- the user can switch between single and partner/multiple use cases, with a nice lean UI and to the point UX -- every pixel must make sense, otherwise it is unimportant.
- sent the updates for review for play and appstore
- about SEO and GEO, lately I bought a new domain and watching its reputation and position in terms of the keywords I am tracking, I have integrated recently Google Analytics to track to keywords better.

- I have analyzed the markets in this field in 24 regions and set goals for them and keywords for them.
It is a rather long process, right now majority of the users are coming to the website and downloading the app via web search and ChatGPT chats, I would like to increase that.

TBH I don't want to spend my time working on the product, rather focusing on the distribution and trying new things -- that's what I would like to do but without having a proper product and onboarding/paywall to communicate with the users' needs, it is not sth good to focus on distribution blindly as well -- the balance is important.

I am curious after having the new release approved, how will the conversion rates change for the single users.

[Semih @its_meseba]

@marclou @RevenueCat @trust_mrr Are you thinking about adding adapty as well, is it possible?
Many people use it instead of RevenueCat