Daniil Agniashvili

[Daniil Agniashvili @NiceDaniilak]

A little detail in my checkers game: chapter cards now open and fold back into place. 🐰♟️

A small touch, but it makes the journey feel a little more alive.

#IndieGame #GameDev #CozyGames #Checkers #iOSGame

[Daniil Agniashvili @NiceDaniilak]

It's raining checkers! ☁️🐰
Light or dark: which side are you choosing? 🤍🖤
#Checkers #CozyGames #IndieGame

[Daniil Agniashvili @NiceDaniilak]

@kuuuuuuuuuua I built a similar app back in October. I didn't think anyone would be interested in it. The images are generated based on daily plans....

[Daniil Agniashvili @NiceDaniilak]

Cloudflare also offers a tool to check whether your website is ready for AI agents

https://t.co/7p5n55SZhR

[Daniil Agniashvili @NiceDaniilak]

I know quite a few of my followers work in SEO, so bookmark this post and share it with your colleagues. You might find it useful.

It turns out that much of today's SEO for AI is just SEO advice from 20 years ago.

Here's what server log studies reveal:

😊 Almost nobody cares about llms.txt. Ahrefs analyzed 137,000 websites and found that 97% of existing llms.txt files received zero requests over a month. Google has officially confirmed that it doesn't use these files for Search, including its AI features.

🌞 JavaScript is a problem for AI crawlers. In an experiment involving 1,062 pages, GPTBot discovered 748 pages through HTML links and not a single one through JavaScript-generated links. This makes SSR and SSG more reliable choices than pure SPAs.

🔴 Markdown offers no proven advantage in AI search. In a study by Dries Buytaert, not a single AI crawler requested Markdown through Content Negotiation. Meanwhile, serving separate .md pages increased crawler traffic by approximately 7%.

So what actually works?

Accessible HTML, clear site structure, server-side rendering, proper internal linking, and high-quality content.

In other words, good old-fashioned SEO.

Cloudflare also offers a tool to check whether your website is ready for AI agents (link in comments)

#SEO #TechnicalSEO #AISEO #AICrawlers #WebDevelopment #LLM

[Daniil Agniashvili @NiceDaniilak]

AI has learned to write code. But who's making sure it actually does what it's supposed to?

It's funny how AI is turning software development into a process of writing documentation and natural-language instructions. Developers used to hate documentation, but at least they understood their own code. Now AI agents write the code, and even developers don't always fully understand how it works.

Testing makes things even more interesting. AI writes the code, unit tests, and E2E tests. But if the same model generates all three, what exactly are those tests verifying? After all, an agent can misunderstand the requirements in exactly the same way when writing both the implementation and the tests.

And that's the paradox: more code is being produced, tests are passing, tickets are getting closed, yet our confidence that the system actually works as intended is decreasing.

The question is no longer just who writes the code, but who takes responsibility for making sure it's correct.

#AI #SoftwareEngineering #AICoding #AgenticAI #SoftwareTesting #LLM

[Daniil Agniashvili @NiceDaniilak]

I'm building Mochi Cam: a little iPhone camera with glossy buttons, pastel shells, and toy-camera nostalgia. ☁️🪐🩷

Cloud, Orbit, or Jelly—which one would you pick?

#MochiCam #Kawaii #BuildInPublic

[Vitaly Kleban @vkleban]

I've built https://t.co/uNwcNbWgkD - fighting game where both players are controlled by Jev models! 🥊

Tell me your fighter name and I will attack you!

[Daniil Agniashvili @NiceDaniilak]

I'm building Mochi Cam: a little iPhone camera with glossy buttons, pastel shells, and toy-camera nostalgia. ☁️🪐🩷

Cloud, Orbit, or Jelly—which one would you pick?

#MochiCam #Kawaii #BuildInPublic

[Daniil Agniashvili @NiceDaniilak]

10. A few months later you have an entire infrastructure wrapped around one LLM.

It works. It also needs constant care, and sometimes it is slower and more expensive than competitors who just called a different model through an API.

[Daniil Agniashvili @NiceDaniilak]

How every startup actually builds its own AI product:

One API call. A system prompt. "Ready for investors."

Months later it's a whole platform around one LLM, sometimes slower and costlier than rivals who just swapped the model.

Save this 10-step guide

[Daniil Agniashvili @NiceDaniilak]

9. Before every release you have to prove the model won't promise payouts past the limit, won't reveal salaries, won't follow instructions from someone else's email, and won't break what already works.

You add automated tests.

[Daniil Agniashvili @NiceDaniilak]

8. Now nobody can tell why an answer is wrong, which tool ran, or where it failed.

You add request tracing, logging, and monitoring.

[Daniil Agniashvili @NiceDaniilak]

7. Caching does cut the bill. Then a user asks if the server is up, and the model replies "yes" from cache. The server is already down.

You set TTLs and decide which facts must always be checked live.

[Daniil Agniashvili @NiceDaniilak]

6. It finally feels safe.

Monday morning, users send 5,000 questions and the token bill hits 999,999 $.

You switch to cheaper models, cut context, and add caching.

[Daniil Agniashvili @NiceDaniilak]

5. A new problem shows up. Someone hides instructions in an email or a doc the model will read.

In one case, prompt injection made an AI refund customers more than 2 million $.

You add injection defenses, input filtering, and checks on dangerous actions.

[Daniil Agniashvili @NiceDaniilak]

4. Now it can restart servers, open tickets, and send messages.

Then an intern asks for the boss's salary and gets it.

You add OAuth, roles, and real access control.

[Daniil Agniashvili @NiceDaniilak]

3. You want a little autonomy.

You add MCP and other tools so the model can talk to outside services.

[Daniil Agniashvili @NiceDaniilak]

2. Then the model invents facts, leaks private data, takes actions it shouldn't, and burns tokens.

So you add RAG: documents become vectors, you search by meaning, and you hand the hits to the model.

[Daniil Agniashvili @NiceDaniilak]

1. Grab GPT or Claude through an API. Add a system prompt.

Everyone decides the product is done. Time to show investors.