17/10/2026 –, Principal
Using simple demos with roles, this talk will show tips and tricks for agentic coding tools. Attendees will leave with a better understanding of context windows, prompting, and Python specific tips.
To use coding agents well, you need an intuition for how they work and what they actually do. Without it, developers fall back on using agents like a more powerful chat window, and then struggle to understand why they get slop and unpredictable results.
The good news: building this intuition is easy. It takes only a little curiosity and a willingness to experiment.
This talk builds that intuition through simple experiments with persona prompts, starting with silly experiments like, "pretend you are a Bjarne Stroustrup" or "pretend you are a goose", and progressing to "pretend you are a Django novice," "pretend you are an expert software architect."
We will solve FizzBuzz in Claude. Suprisingly this simple problem is sufficient to demonstrate some tips and traps when using agents. Along the way, we will demonstrate:
- model non-determinism
- how context windows work
- why context management matters, and how to track it
- session rewinding and forking
- zero-shot vs. multi-shot prompts
- plan mode vs. edit mode
- sub-agents and skills, including parallel exploration
The final demo shows how a default Claude session, without proper setup, produces bad tests and technical debt in Python. We then show how to use skills and sub-agents to produce good tests instead.
Developers will leave knowing how Claude actually works, how to prompt it for good results, and with a sub-agent/skills file they can use to improve their own tests.
The first half should be accessible to all devs, since the sample problem is FizzBuzz, and I will be showing how different prompts produce varying outputs. Zero knowledge of Claude is required, but is helpful.
The final demo will be more suitable for intermediate-advanced devs. It will require deeper knowledge of TDD and testing techniques, and mocking and patching to recognize the differences in the personas and why their outputs are good or bad. Familiarity with Claude (or other agent) will be helpful, since I will show specific features of Claude that affect prompting.
O que as pessoas que participarem podem esperar aprender na sua atividade?:They will learn:
* caveats for using agents and trusting the code quality they produce
* how agents are non-deterministic, and sometimes in surprising ways
* how context windows work, and why they must be managed
* tips on using Claude (plan mode, subagents, skills)
* simple prompting tips and terminology (zero-shot vs. multi-shot)
Ed is an accidental software engineer based in Austin, TX, He's worked in several domains ranging from industrial robots to NLP, and most recently in real estate wire fraud prevention.