Coding with AI

Vibe Engineering - AI theory and practice

Transcript

1. Coding with AI Feb 2026
2. "Evolution, Morpheus. Evolution! Like The Dinosaur. Look Out That Window. You've Had Your Time. The Future Is OUR World, Morpheus. The Future Is OUR Time."
3. Becoming Neo Feb 2026
4. Agenda • Getting intuition for the technology and buzzword • How to maximize value from working with AI and what are the limitations • How to AI
6. LMM - Large Language Model • Large - trained on a huge data set and uses a huge number of parameters • Language - geared toward understanding language • Model - a type of Neural Network
7. So let’s start with the basics where 𝑎∈1...𝐴 is the head number, and 𝑓 is some function like RELU or whatever and the 𝐛s are biases (𝑀 is the attention mask and 𝑑 𝐸is the size of the embedding). the output of layer 𝑙∈1...𝐋, 𝐗𝐥 is ,
8. Ok - so the real basics Preceptron (1957)
9. Network
11. Network https://playground.tensorflow.org
12. embeddings https://projector.tensorflow.org/
13. Discriminative vs. Generative models
17. https://github.com/vectara/hallucination-leaderboard
18. The U attention trap • close attention across an entire window is computationally expensive. • High attention to the start of the input • High attention to the end of the input • Low-to-no attention to the middle https://arxiv.org/pdf/2406. https://www.linkedin.com/pulse/real-limits-ai-attention-u- shaped-truth-context-windows-sergeyev-bk8oc/
21. Consequences • The main algorithm is next word (actually part of word) prediction ◦ What we get is an option • Common things are easy - if you are using common practices, working on areas that have a lot of good example, chances are AI can really push you fast ◦ The corollary is that if you have unique patterns in your code, completely novel area AI will struggle • Getting exactly what you want is hard • Probable != Correct (aka “hellucinations”) • CONTEXT is king ◦ Control size ◦ Manage attention
22. So, how to AI?
23. “Never Send A Human To Do A Machine's Job.“
28. “We Can Never See Past The Choices We Don't Understand.”
30. “I Know Kung-Fu” (or do I)
31. Caveat emptor My way is not the only way, Explore and find what works for your style
32. How I used to work with AI How I work with AI these days Probably not the same
34. Vibe Engineering Design Refactor Review Execute Plan
35. Design • Which models? ◦ Gemini 3 pro in aistudio.google.com • Why? ◦ Search integration (references) ◦ Execute (simulations) ◦ URL content • How? ◦ Bounce ideas ◦ Ask for cons ◦ Clarify ◦ Ask for details ◦ Evaluate options • Almost always needs more refinement since disconnected from actual code
36. Plan • Which models? ◦ Gemini 3 pro (cursor) / claude Opus 4.5 (cursor/claude code) • Why? ◦ Make the high-level plan more actionable • How? ◦ Ask for gaps and challenges ◦ Verify vs. existing codebase ◦ Use multiple models ◦ Decided on scope (defer stuff to todo.md for later) ◦ Align with ADRs ◦ Can use the cc/cursor plan mode • Persist in the memory-bank for tracking and verification
37. Execute • Which models? ◦ Main coding ▪ Claude Opus 4.5 (claude code/ cursor) ▪ Gemini 3 pro (cursor) ◦ Grunt work ▪ Chatgpt 5.2 -seems to have more attention to detail • Why? ◦ As long as we don’t care for costs - better models yield better results ▪ Slow and steady wins the race - but we can augment by running multiple tasks in parallel • How? ◦ Scan through blurb - notice shortcuts and challenge them ◦ Remind models about testing and running tests • Probably worth looking at worktrees (different “branch” for each agent)
38. Review • Which model? ◦ Preferably, a model you didn’t use in coding • Why? ◦ Different perspective • How? ◦ Verify assumptions ◦ Ask model to describe what the code does ◦ Ask models to verify ADRs and plan ◦ Check for todos ◦ Update-docs - (skill/rules) to look at all the changes and how they affect our memory-bank • Map out what’s left - think if you need a new prompt/agent or can continue conversation
40. Refactor • Which model? ◦ Basically a new cycle with all the mods • Why? ◦ Models hallucinate and cheat • How? ◦ Review all the changes by yourself ◦ Ask models to produce documentation (sequences, block diagrams etc.) ◦ Ask models to explain how things work based on code ◦ Ask models to find gaps, todos, “for now” , “fix later”s etc. • Maybe not each cycle but a must
41. Vibe engineering Design Refactor Review Execute Plan
42. How does that fit to LLMs - Breaking into skills and phases - Controls the context size - ADRs and docs - Provide long-term memory - Breaking work into phases - Control contexts size - Manages attention drift - Checkpointing for hallucinations - Review phase, Refactor phase - Manage hallucination - Manage overload , context switch and human-machine disconnect - Skills for security, testing etc - Manage AI risks - Todo.md - Intentionally control scope size
44. Not the only game in town • https://github.com/steveyegge/beads • https://www.vibekanban.com/ • https://github.com/romiluz13/cc10x • https://github.com/SuperClaude-Org/SuperClaude_Framework
46. Skills, MCPs, Rules, Sub-Agents oh my • Rules - e.g. Claude.md, agents.md ◦ always in context (cursor let’s you scope them with glob) ◦ claude let’s you scope them by folder • skills/commands - prompt + scripts + templates - loaded as needed • Mcp - remove or local code execution of ready-made commands
47. Building a skill is Promptgraming
48. Maybe turn a skill to an agent
50. Plan
51. "I Don't Know The Future. I Didn't Come Here To Tell You How This Is Going To End. I Came Here To Tell You How It's Going To Begin.