AI & Enginnering

Transcript

1. AI & Engineering Kaz Yoshikawa Jan 2026 kaz.yoshikawa@gmail.com
2. About me
3. Kaz Yoshikawa • Electricwoods LLC / Digital Lynx Systems Inc. / Slow-Walkers Inc. • e-mail: kaz@digitallynx.com, kaz.yoshikawa@gmail.com • X: @codelynx1 • Working History • Yumemi, Pixiv (Part-timer) • Adobe Systems (Tokyo) • Lionbridge (Tokyo) • Quark (Tokyo / Denver) • Hummingbird Communications (Mt. View, USA) • Fact International (Vancouver, Canada) • Perle Systems (Toronto, Canada), etc.
4. History
5. AI & Engineering
6. AI coding adaptation • Claude Code 85%, Codex 10%, Gemini 5% • NO MCP • NO Skills • Local LLM – Qwen3 (experimental)
7. Do AI do coding? • Before Chat GPT 3.5 • nah, not gonna happen • After Chat GPT 3.5 • may, may be... • After Claude Code • Unfortunately, sure
8. Chat GPT 3.5 era • Simple Prompt • write swift code to extract data from JSON • copy & paste between Xcode • build error – hard to fi x them • copy & paste between Xcode or fi x by human • Work with Apps extension • Chat GPT see your working fi le but not rest of other fi les
9. Claude Code era • CLI tools • Claude Code sees, writes fi les (agentic) • Claude Code knows how to grep, fi nd, xcodebuild and many other commands • Claude Code knows how to debug by observing stdout or stderr • add some log, run, see log, fi x, test, remove log
10. Claude Code Codex Gemini • Claude Code • Good for Swift coding, an engineer must be in it • If the code not working, debug it • Codex • Good for Swift coding, but may not fi x its own build errors • A scientist must be in it, not an engineer • If the code not working, focus reasoning • Gemeni • Good with mix with web search • Asking just review start fi xing code, "oh no, just review, pls" • Claude code was like that before
11. Tips for AI Coding
12. NO / Less Vibe Coding • Make more speci fi c what to develop • Only good for YouTube or Live Demo for surprise • You need to know what you are going to develop rather than expect serendipity
13. More speci fi c • Make a concrete plan before coding • Purpose of the products, what's for? • What's input? What's output? • What's the mechanism behind the scene?
14. Don't be too speci fi c • Give some room for AI for implementation • If there are well know better solutions, too speci fi c instruction overwrites it • Your too speci fi c plan may leads to unexpected goals
15. Work on small • Try break large task into small task • This is traditional engineering rule of thumb • If this cannot be done well, products may not be well designed
16. Plan Mode • Make a concrete plan before coding • Chat with AI about what is the purpose • magic word • Analyse my intent? • Any ambitious? • Ask AI data fl ow for an action or transaction • Give example actions and let AI describe behind the scene Claude code
17. Get Review by Codex • Once a plan is complete and ready to code: • Ask Codex, Gemini or another Claude Code to review the plan • Claude Code: Yes Man // • Gemini: Yes Man • Codex: Very Picky • Copy & Paste Feed back and update plan until clear • Be careful, sometime it end up with your unexpected way
18. Attention Control • AI my have forgotten Features working before • Hallucination e ff ect and AI pretends it knows • Let AI read documentation and related code • > read all documentations and code related to auth
19. Ask AI to write test code • Claude Code id good at interact with stdout or testable code • Make your code more testable (no user interaction) then AI can test well • Ask AI edge case, story based case, AI can test well
20. Ask AI to write POC • If you work on not small software component, ask AI to write POC (proof of concept) • You can try and error as much as you want with AI • If it worked, then bring the code to main code line • If it is not working, drop the branch
21. UI debugging • UI is hard to debug for AI • Ask AI to add some log for UI related code • We can copy paste log to UI, to fi nd UI related bugs • Resent AIs can see or analyze screenshots • Mac, we can use OCR from preview
22. AI can code with overwhelmed Spec • Adobe PDF Speci fi cation 100 • Microsoft O ff i ce Format Speci fi cation • Word, Excel, Power Point etc • RFC
23. Academic Paper Polar Stroking: New Theory and Methods for Stroking Paths MARK J. KILGARD, NVIDIA Fig. 1. Polar stroking samples: A cubic Bézier segment with a cusp rendered properly with polar stroking while uniform parametric tessellation has no cusp, both using 134 triangles; B polar stroking improves the facet angles distribution compared to uniform tessellation, both using 126 triangles; C arc length texturing; D ellipse drawn as just 2 conic segments, one external; E complex cubic Bézier path (5,031 path commands, 29,058 scalar path coordinates) with cumulative arc length texturing; F centripetal Catmull-Rom spline. Stroking and !lling are the two basic rendering operations on paths in vector graphics. The theory of !lling a path is well-understood in terms of contour integrals and winding numbers, but when path rendering standards specify stroking, they resort to the analogy of painting pixels with a brush that traces the outline of the path. This means important standards such as PDF, SVG, and PostScript lack a rigorous way to say what samples are inside or outside a stroked path. Our work !lls this gap with a principled theory of stroking. Guided by our theory, we develop a novel polar stroking method to render ACM Reference Format: Mark J. Kilgard. 2020. Polar Stroking: New Theory and Methods for Stroking Paths. ACM Trans. Graph. 39, 4, Article 145 (July 2020), 15 pages. https: //doi.org/10.1145/3386569.3392458 1 INTRODUCTION Vector graphics standards such as PDF [Adobe Systems 2008], SVG [SVG Working Group 2011], PostScript [Adobe Systems 1985], PCL 308v3 [cs.GR] 31 Oct 2020
24. AI can code what you don't know much about • I learned AWS with Umedy, but still it is hard • IAM, Policy • S3, Life cycle, STS • SES, SNS • AI do code what I have very few knowledge about // codelynx case
25. AI may not design more than what you are capable of • Designing Storage feature in S3 • Rename folder feature • What!? S3 cannot rename intermediate directory name? • why you didn't tell me!? // codelynx case
26. Examples // codelynx case
27. Swiftlet A modern Swift web framework with file-based routing and declarative HTML generation Swift Swift 6.0+ 6.0+ Platform Platform macOS | Linux macOS | Linux License License MIT MIT Live Demo: http://swiftlet.eastlynx.com:8080/ Warning Development Status: Swiftlets is under active development and is NOT ready for production use. APIs may change, and features may be incomplete. We welcome feedback and contributions to help shape the future of Swift on the web! Swiftlets is a lightweight, Swift-based web framework that brings the simplicity of file-based routing and the power of type-safe HTML generation to server-side Swift development. Inspired by modern web frameworks, it offers a unique approach where each route is an independent executable module. File-Based Routing - Your file structure defines your routes ( .webbin files) Declarative HTML DSL - SwiftUI-like syntax for type-safe HTML generation SwiftUI-Style API - Property wrappers ( @Query , @Cookie , @Environment ) for easy data access Zero Configuration - No complex routing tables or configuration files Security First - Source files stay outside the web root, MD5 integrity checks Hot Reload - Automatic compilation and reloading during development Swiftlets What is Swiftlets? Key Features README Inspired by Ignite // codelynx has very few knowledge about server side
28. Di ff a
29. Kronoth (Server/Browser) • S3 base peta scale version control system • Git like un-destructive content management system • Swift code base
30. Hallucination for engineers
31. Can we make with AI • Photoshop, Illustrator, InDesign • Word, Excel, Power Point • EC Site • Government / Business System • Embed System
32. Get Real...
33. How to live as an engineer with AI?
34. Written by Claude Code
35. Demographic of Engineers • Junior class engineers may extinct fi rst • Professional class engineers may vanishing • Senior class engineers retire sooner or later
36. Engineers • Programmers • Architects • System Engineer • Tester • DevOps • Tooling Engineer
37. Be directive engineer • From writing code to directing systems. • The engineer de fi nes the problem; AI executes. • Stop coding. Start engineering. • Intent over implementation. • Good output needs good orders.
38. Who knows the future? • Coding skills still helps directing AI coding • Prompt may not be human language someday • Hang on!
39. Thank you Codelynx / Kaz Yoshikawa kaz.yoshikawa@gmail.com