Kuba Boczek

Engineering leader with a backend/.NET & Azure background. Building data-heavy systems, leading teams, and actively using AI to develop solutions.

[Rohan Paul @rohanpaul_ai]

Claude Code creator Boris Cherny ( @bcherny )

"For people who aren’t building agentic products but are using Claude Code, every 6 months, delete your claude.md file, delete your skills, and delete your hooks. Then see what the model does. It might surprise you.

For Opus 5, we strongly recommend trying to delete all of these things because the model may no longer need the extensive instructions that were necessary for previous models."

- at Y Combinator Startup School 2026.

----

From "Y Combinator" YouTube channel, (full video link in comment)

[Kuba Boczek @jboczekk]

By lowering the entry barrier, the agent made our new analytics easier to use and directly increased adoption.

That’s when Copilot Studio started delivering real business value.

[Kuba Boczek @jboczekk]

To be honest, I wasn’t a fan of Microsoft Copilot Studio at the beginning.

The setup was complicated, the results were disappointing, and the whole product felt like a false start.

Then MS introduced the new Copilot Studio in June, so I gave it another try.

[Kuba Boczek @jboczekk]

That was enough.

Unlike the prev version, the new one was able to use the skill correctly.

The agent can now answer questions about our data, generate the correct queries, and give users access to the results.

But the biggest win isn’t query generation itself.

[Kuba Boczek @jboczekk]

Copilot lasted me 17 active days.

Codex still had plenty of headroom, despite broadly similar usage.

Coding-agent comparisons tend to focus on model quality. How much real work fits within the subscription may matter just as much.

[Kuba Boczek @jboczekk]

@OpenAIDevs Nice! No more manually coping paths to other projects 💪

[Kuba Boczek @jboczekk]

The more interesting result came from analyzing my July session history: how often did each agent actually use RTK?

- Codex: 825/905 - 91.2%
- Copilot CLI: 284/579 - 49.1%

That gap closely matched what I had been observing in daily use (this was not a controlled benchmark)

[Kuba Boczek @jboczekk]

I’ve been testing an optimization tool RTK that removed ~1.8M tokens of terminal output before it entered my coding agents’ context.

The direct cost saving is only a few dollars.

The more interesting result: Codex and Copilot CLI followed the same instruction very differently.

[Kuba Boczek @jboczekk]

RTK reduces shell output before the agent sees it.

Across 17 days:

- ~2k commands
- 3.5M input tokens
- ~1.8M tokens removed

Pricing saved tokens against current model rates gives a direct saving of only a few dollars. Whether that's meaningful depends on your scale and usage.

[Kuba Boczek @jboczekk]

@matt_gray_ High output management is something I have to read in near future. Thanks!

[Ben Meer @SystemSunday]

How to budget.

The 50-30-20 Rule:

[Kuba Boczek @jboczekk]

@ItsKieranDrew Agree. And it's good to learn, to not be afraid of using this sentence.

[Kuba Boczek @jboczekk]

@Codie_Sanchez That's right. The best days for me are those when I wake up before 6am, take cold shower and complete my most important tasks before 9 am.

[Kuba Boczek @jboczekk]

@RyanMSchnitzler Plan tasks for next week

[Kuba Boczek @jboczekk]

@FitFounder It looks like a great plan.

[Kuba Boczek @jboczekk]

@themattmic Junk food, but all are highly destructive.

[Kuba Boczek @jboczekk]

@FitFounder Agree.
When I'm trying to work after 8pm it took me usually twice as much time to complete it compared to that same work done early like 5/6 am.

[Kuba Boczek @jboczekk]

@adcock_brett Great questions. I will use them to build my side project.

[Kuba Boczek @jboczekk]

@gregisenberg Great tips.

I try to use AI wherever I can. It saves me a lot of time.