Coding with Windsurf

Optimizing for Cost and Productivity

Hello, I'm Alper 👋

• Explorer of AI
• Builder of Things
• Father of Three

Self-employed Consultant & Developer

Engineering Lead @ Genius Sports

@alperortac

What is Windsurf?

Windsurf is a VS Code fork with built-in AI capabilities that assist with coding

👇 Me while I'm editing this slide 👇

Why not ... ?

Cursor

GitHub Copilot

Claude Code

JetBrains AI

Ampcode

Antigravity

Kilocode

Kiro

Opencode

Codex

Personal preference. All of these are great.

Costs

Additional roll-over credits

+250 for $10

Token Count Credits

Comparison: Opus 4.5

Tokens per Dollar: Opus 4.5

Which Model?

It depends...

• Changes over time
• Complexity of task
• Promo models

Me currently:

Opus 4.5 (4x), Sonnet 4.5 (2x) and SWE 1.5 (free)

Rules

AGENTS.md

						
							+ # Rules
							+ Use types, never `any`
							
					

In your project root directory

Browser Verification

MCP or CLI to interact with a real browser

Has access to browser console, network tab, lighthouse audits, etc.

Choose Wisely

• Playwright CLI (best to avoid context rot)
• Playwright MCP
• Chrome Devtools MCP (most features)

Rules #2

AGENTS.md

						
							# Rules
							Use types, never `any`
							+ Use playwright-cli to verify UI changes
							
					

Backwards Compatibility

Rules #3

AGENTS.md

						
							# Rules
							Use types, never `any`
							Use playwright-cli to verify UI changes
							+ Never add backwards compatibility without explicit approval
							
					

Writing Tests

Rules #4

AGENTS.md

						
							# Rules
							Use types, never `any`
							Use playwright-cli to verify UI changes
							Never add backwards compatibility without explicit approval
							+ Never loosen test assertions just to make them pass
							+ Failing real tests is better than fake passing tests
							+ You can declare tests as ready even if they are failing
							
					

Context Window

Most models have a 200k token context window (a few have 1M)

Once >50% full, performance degrades significantly

Distractions that fill context fast:

Context Curation

🧹 What can we do to keep the context clean? 🧹

Fast Context

Why it's great

• It's a subagent → context stays clean
• Extremely fast ⚡

Rules #5

AGENTS.md

						
							# Rules
							Use types, never `any`
							Use playwright-cli to verify UI changes
							Never add backwards compatibility without explicit approval
							Never loosen test assertions just to make them pass
							Failing real tests is better than fake passing tests
							You can declare tests as ready even if they are failing
							+ Always use Fast Context when searching for file contents
							
					

Interviews

Chat asks for clarifications:

1. Less guessing
2. Better requirements

Interviews #2

Every Opus prompt costs 4 credits - no matter the token count

Interviews #3

The Ask tool fills the prompt token context with requirement gathering

Each ask round-trip adds context without spending credits

Rules #6

AGENTS.md

						
							# Rules
							Use types, never `any`
							Use playwright-cli to verify UI changes
							Never add backwards compatibility without explicit approval
							Never loosen test assertions just to make them pass
							Failing real tests is better than fake passing tests
							You can declare tests as ready even if they are failing
							Always use Fast Context when searching for file contents
							+ Ask before, when blocked mid-task and to confirm unknowns
							+ Never guess, never assume or improvise unagreed solutions
							+ Only I can declare a task complete
							
					

Planning

Communicate intent, not implementation details

Simple markdown files and/or plan mode are sufficient

Skill Tree

Apart from Rules there are:

• Memories
• Skills
• MCP's
• Workflows (Commands)

I use them only sparingly because I prefer progressive improvements.

Troubleshooting

Let AI show you its system prompt

Why didnt you follow the instructions about legacy code?

Which part of the rules make you write tests like that?

Curiosity > Judgement

Little Helpers

• Autocomplete
• Commit Message Generation 👉
• Lifeguard reviews code for issues

Magic Wand: If I could have one new Windsurf feature

Subagents with isolated context are the missing piece

Each subagent has its own context to avoid pollution

Next steps

Swiss-cheese model and spec-driven development

source: https://www.latent.space/p/reviews-dead

AI Stack

Building a community for AI enthusiasts

https://aistack.to

Thanks!

Questions?

https://x.com/alperortac