datafrosch.fun/slides/code-anything Either use codex or claude code, otherwise: Download Opencode Download Python

Hands-on (5–10 min)

  1. Open ChatGPT, Claude or DeepSeek — chat apps that make interactive artifacts.
  2. Prompt it to “make a snake game”.
  3. Have fun with the prompt — make it poop rainbows, make it a space version.
  4. Play it for a bit, then paste a screenshot & the prompt on the shared board.
Plain snake game (GPT) GPT — plain
Plain snake game (DeepSeek) DeepSeek — plain
Desert snake (Claude Sonnet) Claude Sonnet — desert
Space snake (Claude Opus) Claude Opus — space
Tea-party snake (Claude Haiku) Claude Haiku — tea party
Workshop

How to code
anything

Dataharvest 2026

Trainer

Ada
Homolova

Cooks with Deepseek V4 Pro

datafrosch.fun
Trainer

Johan
Schuijt

Almost a language model himself.

resolve.works

What have we coded
with LLMs?

🌼 This presentation 🌼

What happens when we prompt for 🐍?

  • The model knows what a “snake game” is — thanks to its training data.
  • It knows how to make interactive artifacts (JavaScript).

What happens when we prompt for 🐍?

  • But: it will make a lot of arbitrary choices on its own.
    • How does it look?
    • How does it behave?
    • Do walls end the game?
    • Are there obstacles?
    • Does the snake have eyes?

We don’t want this 🙅‍♀️

Topics

  1. Prompt
  2. Context
  3. Local agents
  4. Plan → Implement → Review
One

Prompt

The optimal prompt sits between vague and overly prescriptive

Effective context engineering for AI agents

Prompt example

✍️🤖

Prompt example, take two
Two

Context

Making sure the model gets enough information to perform its task well.

Context:

  • Documentation
  • Methodology
  • Reports
  • Examples
  • Tools
Context vs prompt

Effective context engineering for AI agents

No context = just training data = more randomness
Context benchmark

FACTS Benchmark Suite Leaderboard

WHY?

Mechanistic reality

Good practices for context building

Discuss

Talk through the problem with the model first.

Ask me questions

Make the model probe your assumptions.

🐍

Prompt

I want to build a simple snake game. What are the steps I need to think about from the core mechanics up to the visual design?

Answer

1. Core mechanics — represent the snake, move it, handle arrow keys.
2. Game world — define the grid; walls or wrap-around.
3. Food & growth — spawn food, eat it, grow the snake.
4. Collisions — game over on wall or self-collision.
5. Scoring — points, restart, maybe a high score.
6. Visual design — colours, polish, sound.

Each of these steps might need different context.

The smaller the model, the more important it is to divide the work well.

What is a "smaller" model?

Model size comparison
DeepSeek V4
Context — takeaway
  • Context is what the model can see right now.
  • Build the context through a conversation with the model and external sources.
  • Split big tasks and provide only the relevant context.
  • The smaller the model, the smaller the chunks it can handle.
Three

Local
agents

Free the robot

Hands-on: what is an agent? (5 min)

  • Ask the agent what an agent is.
  • Ask the agent what it can do.
  • Ask the agent what tools it has.

Agent
=

a model with the ability to execute tools.

Local Agent
=

uses tools that can read and write files to your computer.

Local agents — takeaway
  • An agent = a model + tools that can act on your computer.
  • Most local agents can access your files and execute code
    ❗ be aware of the risks.
  • Collection of tools is called a harness.
Four
Plan Implement Review

Hands-on (15 min)

Make a scraper for books to scrape.

  1. Plan — discuss what needs to happen with the agent to minimise random behaviour. Use "plan mode".
  2. Implement — grab the plan and have the agent implement it. Use "build mode".
  3. Review the output.

Iterate (for larger projects)

  • Git
  • Automated testing
Plan · Implement · Review — takeaway
  • Plan first — discuss before any code gets written.
  • Implement second — let the agent run; you steer.
  • Review always — never trust output you haven't read.
  • Iterate — for bigger work, lean on tests and git.

Hands-on

Let us help!

Ask us for an API key for DeepSeek V4 Pro for more ambitious projects
  • Make a dashboard from the scraped data
  • (Start) work(ing) on your own project
Join 🪷The Pond🪷!

datafrosch.fun