# AI Coding Holiday Coding *March 6, 2025*
## Agenda - What is Holiday Coding? - Inspiration - Dive in! - Recap and Q&A

Holiday Coding

Holiday coding is improving production systems by applying understanding through effort. It is a playful way of learning on the job in off-time.

Playful
Effort
Production

!Holiday Coding

Defined as what it is not.

No Exams
No Expectations
No straight path
## Inspiration and Hooks These are lessons learned, not laws. Ideas to spark your interest.

Inspirational Blogpost

Link here

Remarkable Tool

Link here

Great Podcast Episode

Link here

Dumbing it down / Catching up.
LLMs are fundamentally massively trained neural nets ➡ Brains
## Types of brains: - Doing clear repetitive stuff. - Knowing things - Thinking things - Narrow Minded - Open Minded
## Speed of brains: - Slow - Fast - Consistent
## You can make brains collaborate - Reason first, Come up with a plan - Execute plan. - Check plan.
- Provide context to - Write a prompt - That generates prompts - to execute tasks - that generate context - repeat hence: Prompt engineering
## Context = King - By giving it context manually - By finding the right documents (RAG) - By steering and summarizing.
Model Context Size Speed Reasoning
03-mini-high 2048 tokens Fast No
4o 8192 tokens Medium Yes
4o schedules 8192 tokens Medium Yes
o1 4096 tokens Fast No
o3 16384 tokens Slow Yes
## AI Coding - Do things - Don't do Things - Explore & Learn
## Game Changing
## Doing Things Better / at all / faster - Documentation - Tests - Coverage of features.
## Documentation Better for humans and AI! The Bar is raised 100x
## Downstream Documentation Chain Help AIs understand your software. E.g. Document your API so that the people using your API can use LLMs to implement.
## Quality Assurance Due to the increase agency of LLMs it is irresponsible not to have a good test coverage. Either you have/generate them, and leverage the speed of LLMs. Or you don't, and you are slow and will lose out in the market.
## Architecture will make the long term difference Architect with a goal in mind.
## Software can Evolve much faster Integration tests will be key to speed.
## Code to infrastructure is a continuum Especially using cloud services. CDK - Pulumi - Terraform / Cloudformation will be driver of speed.
Infra as code is no longer optional if you want to stay competitive
Security, speed, flexibility, costs. If you don't specify you get "the average".
## Pitfalls
It chooses the most used way. Crowds can be stupid.
It helps you, until it doesn't. If you're in way over your head, you might look at a wall instead of a hill.
It's brain is bigger than yours. The steps it takes might be bigger than you are used to. Not unlinke senior vs junior engineer.
It's can get very expensive very fast. Fuck around and Find out now burns cash.
## Lessons Learned
It's remarkably good at coming up with a plan.
Context is as important for LLMs as it is for humans.
Asking the system to summarize things, seems to solidify context.
Following a workflow that resembles your own, keeps you aligned.
It actually is much smarter than you think. If you give it the right context
It can reason across languages, frameworks in a single repository.
Scary!
Based on a spec, it can write a test suite, implementation and iterate if you give it freedom to execute.
It can make rediculous mistakes in logic.
## Holiday Wanderings

Claude Code

Get it here
## Try - Make a plan for your next feature - Document something. - Generate tests. - Make something 100% complete. - Ask it to explain your own code.
# Your Turn
## Discussion Time - What inspired you? - Where did you wander?
# Thank You!