Hurricane Hackathon — Workshop I
Vibe Engineering with Claude Code
Saturday, May 2, 2026 — University of Tulsa
Dallas Elleman
“AI won’t replace humans — but humans with AI will replace humans without AI.”
— Karim Lakhani, Harvard Business School professor, August 2023
“You’re not going to be hacked by an adversarial AI; You’re going to be hacked by an adversary who knows how to use AI.”
— Anthropic CEO Dario Amodei (probably)
“You’re not going to get screwed by a malfunctioning AI torching your codebase; you’re going to be screwed by bad engineering that allows a malfunctioning AI to torch your codebase.”
— anonymous Amazon engineer
Dallas Elleman
2000Austin TX — Aspiring Rock Star
2002–12Drifting → Fatherhood → Manual labor
2012–16TCC — AS EE / Physics / Math
2014NASA JPL — NDEAA Intern & Dodgeball Captain (champion)
2015Symbiotic Aquaponic — Embedded Systems Dev
2016–19UTulsa — BS Eng Physics / Robotics
2019–21XWorks — Robotics/UAV Engineer
2020Dept of Energy — PCAST
2021–23UTulsa — MS Computer Science
2023–nowUTulsa — Cyber Fellow, PhD Cyber Studies
2025–nowDais — AI Product Dev
Spr 2026UTulsa — Teaching CYB-4203/6203 Secure & Trustworthy AI
Agentic Eras Timeline
Era 1
The Chatbot Copy-Paste Era
mid 2022 – early 2023
Jun 2022GitHub Copilot GA— First commercial AI pair programmer, powered by OpenAI Codex
Oct 2022Copilot CLI prototyped— GitHub Next Oxford offsite seeds the CLI-agent idea
Nov 2022ChatGPT launches— Kicks off the browser copy-paste coding era
Era 2
The LLM IDE Era
2023 – 2024
Mar 2023GPT-4 + Copilot Chat— Conversation moves into the IDE
Mid 2023Code Interpreter— First mainstream agentic coding sandbox
Mar 2024Devin launches— SWE-Bench jumps from ~2% to 13.86%
Jun 2024Claude 3.5 Sonnet— Becomes dominant backend for Cursor and coding tools
Sep 2024OpenAI o1— Reasoning models change how LLMs handle debugging
Oct 2024Claude 3.5 v2 + computer use— Browser and terminal control
Nov 2024Cursor rises + MCP announced— AI-native editor proves categorically better than plugins
Era 3
The LLM CLI Era
late 2024 – 2025
Jan 2025The DeepSeek Moment— R1 matches o1 at $5.6M; Nvidia loses ~$600B in one day
Feb 2025Claude Code launches— CLI-first agentic coding sets the template
Feb 2025Copilot Agent Mode— Autonomous multi-file editing in VS Code
May 2025OpenAI Codex + Google Jules— All three major labs now have flagship coding agents
Jul 2025Kimi K2 + Qwen3-Coder— Open-weight models match frontier performance
Mid 2025Vibe coding goes mainstream— Lovable hits $100M ARR in 8 months
Jul 2025Claude 4 + GPT-5— New frontier models arrive simultaneously
Era 4
The Claw Era
2025 – present
Nov 2025Kimi K2 Thinking— 200+ sequential autonomous tool calls without losing the thread
Nov 2025OpenClaw launched— Personal AI agent that lives on your entire machine
Jan 2026OpenClaw goes viral— 60k stars in 72 hours, 300k+ total; defines a new category
Jan 2026Kimi K2.5 Agent Swarm— 100 parallel sub-agents; adopted by Cursor Composer 2.0
Early 2026MCP becomes universal— Standard across all major IDEs and coding agents
2026SWE-Bench exceeds 80%— Up from ~14% when Devin launched two years ago
2026Anthropic Mythos / Glasswing— Discovering 22-year-old zero-days in OpenBSD
Concepts
Vibe Engineering
Human Input:
Knowledge
Structure
Supervision
LLM Output:
Mistakes
LLMs as Pure (albeit probabilistic) Functions
Tokens In
→
Semi-opaque Box
The Context Window
→
Tokens Out
“Some things are within our power, while others are not.”
— Epictetus, ca. ~100 AD
Concernradar
Influenceaims
Controlactions
Concernradar
Circle of Concern
Aspects of LLMs we have no control over
- Inherent capabilities and limitations
- Non-determinism and hallucination
- Unpredictable effects of widespread adoption
Influenceaims
Circle of Influence
Aspects of LLMs we can influence, but not absolutely control
- Model output quality, accuracy
- Reliability, maintainability, security, readability of generated code
- Which labs win (shout out to Anthropic’s Project Glasswing)
Controlactions
Circle of Control
What we should focus on
“The only thing that improves the quality of your outputs is the quality of your inputs.”
— Dex Horthy, HumanLayer
“The job of Software Engineers is now to be Locomotive Engineers and keep the LLM train on the tracks.”
— Geoffrey Huntley
The Context Window
The Context Window
Context Engineering
“The science of optimizing context window input and thereby optimizing LLM outputs.”
- Context Retrieval and Generation
- Context Processing
- Context Management
- Retrieval-Augmented Generation (RAG)
- Memory Systems
- Tool-Integrated Reasoning
- Multi-Agent Systems
Research → Plan → Implement
Spec Driven Development
“LLMs can read, write, understand code very effectively, and are improving continuously. The idea is that rather than writing precise code, we should focus on writing precise specification documents that our LLM assistants can read and update — Specs are the ground truth that can be translated into any desired code language by LLMs.”— lots of people
Spec Driven Development
“In an era where AI is transforming software development, the most valuable skill isn’t writing code any more — it’s communicating intent with precision.”— lots of people
llm-dev
Claude Code plugin for spec-driven, transcript-tracked development.
github.com/DallasElleman/llm-dev
Vibe Engineering with Claude Code
1CLAUDE.mdProject instructions & conventions
2PermissionsRead, Edit, Bash — granular control
3Plan ModeThink before acting
4CheckpointsRevert to any prior state
5SkillsReusable prompt libraries
6HooksPreTool, PostTool, Notification
7MCPConnect external tools & services
8PluginsSkills, Hooks, Docker, pytest…
9ContextFiles, history, tools, rules
10Slash Commands/review, /commit, custom…
11Compaction100k tokens → 10k tokens
12SubagentsParallel task delegation
Categories
Security
Context Mgmt
RPI
Claude Code vs. Everyone
Simon Willison’s
Ralph Wiggum Loop
How to Vibe Engineer in 2026
Vibe coding is dangerous and sloppy. Vibe engineering is careful, methodical, and effective. You only get better with practice. Now is the time to practice.
Read the docs. Tell Claude to do the same.
Brainstorm first. Forethought beats firepower.
Plan — but not too big. Your plan changes when it hits reality.
Dispatch agents. Subagents for research, parallel tasks, scoped work.
Shout out to Fazil Raja — check his guide:
github.com/fazilraja/how-to-vibe-engineer-2026
github.com/fazilraja/how-to-vibe-engineer-2026
Thank You
“See what you can come up with.”
Free week of Claude Code
claude.ai/referral/c5aDmiMipg