How I code with AI changed a lot
🔗 **Source:** [How I code with AI changed a lot](https://www.youtube.com/watch?v=xJaMTo2YgO8)
> [!note]+ Document Notes
> tags: [ai-development, coding-workflow, developer-tools, gpt-4, remote-development]
> author: [MISSING]
> source: YouTube
> reference: https://www.youtube.com/watch?v=xJaMTo2YgO8
> date created: {{date:YYYY-MM-DD}} {{time:HH:mm}}
>
> # 🏷️ Title: How I code with AI changed a lot
>
<!-- deep-read:sections start -->
## 🧭 Table of Contents
- [📝 Summary](#-summary)
- [📖 Main Content](#-main-content)
- [🗒️ Notes](#️-notes)
- [📚 References](#-references)
- [🗣️ Paraphrase](#️-paraphrase)
- [📰 Newsletter](#-newsletter)
- [✅ Action Items](#-action-items)
- [🧪 Hypotheses](#-hypotheses)
- [📊 Methodology](#-methodology)
- [❓ Questions](#-questions)
- [🔮 Future Research](#-future-research)
- [💡 Implications](#-implications)
- [🧠 Learning Styles Assessment](#-learning-styles-assessment)
- [📊 Accessibility Assessment](#-accessibility-assessment)
- [🌍 DEIB Assessment](#-deib-assessment)
- [✅ Logical Validity](#-logical-validity)
- [🛡️ DARVO Analysis](#-darvo-analysis)
- [🎯 Conclusion](#-conclusion)
## 📝 Summary
- Creator has completely changed their AI coding workflow in 5 months, moving away from Cursor's plan mode to GPT-4.5 with simpler tools
- Built "Lakebed" - a new full-stack framework/cloud platform - demonstrating the evolved workflow
- Key shifts: GPT-4.5 over Claude, Codex/T3 Code over Cursor, remote development, single-threaded work, minimal agent config files
- Focus on context management, reading AI outputs thoroughly, and steering models through conversation rather than complex prompts
- Remote development setup enables coding from anywhere without losing work when disconnecting
## 📖 Main Content
The creator dramatically shifted their AI coding approach while building "Lakebed," a new full-stack framework designed as a "shitty cloud for shitty apps" to simplify agent-driven development.
### Key Technology Changes
**Models**: Switched almost entirely to GPT-4.5, abandoning Claude models despite previously preferring them. The $200/month plan provides effectively unlimited inference, especially with the 10x multiplier from attending an OpenAI event.
**Development Environment**: Moved from Cursor to Codex CLI with T3 Code or the Codex app. T3 Code is an open-source project (not T3 Chat) that manages multiple AI harnesses in a unified interface.
**Remote Development**: Heavy emphasis on coding on remote machines (like a Mac Mini on local network) while controlling from laptop. This enables closing the laptop mid-session and resuming work later without interruption.
### Workflow Philosophy Changes
**Context Management**: Instead of complex agent configuration files, uses minimal `.agent.md` files written as "letters to the agent" explaining project philosophy and terminology rather than technical rules.
**Single-Threading**: Abandoned parallel work trees in favor of one thread per task on main branch. Creates 20+ threads per day but completes each fully before moving to next.
**Conversation Over Configuration**: Emphasizes reading AI outputs thoroughly and steering through natural conversation rather than elaborate prompts. Most prompts are 2 sentences or less.
### Technical Implementation
Built Lakebed (later renamed from "Span") through voice-to-text planning sessions, starting with high-level goals rather than implementation details. Used HTML output for better-formatted plans and extensive screenshot-based feedback.
The agent configuration focuses on project philosophy and term definitions (distinguishing "me/us" building the tool vs "developers" using it vs "agents" the developers use) rather than file paths or technical constraints.
## 🗒️ Notes
### Model Selection and Usage
- GPT-4.5 now preferred over Claude models for coding tasks
- $200/month plan with 10x event multiplier provides effectively unlimited inference
- Fast mode usage increased due to generous limits on Codex plan
- Occasionally uses Claude for quick landing pages but 4.5 for serious development
### Development Environment Evolution
- Cursor largely abandoned except for cursor.glass UI (which has reliability issues)
- Codex CLI with T3 Code provides superior experience to Cursor plan mode
- T3 Code supports multiple harnesses (Cursor, Claude, OpenCode) in unified interface
- Codex app best for most users, but T3 Code preferred for stability and remote features
### Remote Development Infrastructure
- Mac Mini on local network controlled via screen share and remote apps
- Codex remote experience plagued with issues: disappearing model picker, 30-second delays, sticky keys
- T3 Code remote hosting much more reliable, works via browser with full speed
- Tailscale integration enables secure remote access across networks
- Example: Developer codes entirely from Android tablet using T3 Code server on Replit
### Context Management Strategies
- Clone relevant codebases for reference during development sessions
- Voice-to-text for initial planning dumps creates better prompts
- Minimal agent.md files focus on philosophy, not technical specifications
- Glossary of terms helps distinguish roles (builder vs user vs agent vs developer)
- HTML-formatted plans much easier to review than text
### Workflow Patterns
- Single-threaded development: complete one task fully before starting next
- New thread for each distinct concern to avoid context contamination
- Read AI outputs thoroughly before focusing on code - text often more important
- Steer models through conversation rather than complex initial prompts
- Screenshot tools with annotation crucial for UI feedback
### Anti-Patterns Identified
- SSH terminal coding with AI severely limited (no image paste, complex tmux setups)
- Plan mode too restrictive compared to natural conversation
- Parallel work trees create too much context overhead to track
- Over-engineered agent configurations less effective than simple steering
- Skipping AI text output in favor of just reading code
## 📚 References
- [Lakebed Framework](https://github.com/author/lakebed) - The project built demonstrating this workflow
- [T3 Code](https://github.com/t3-oss/t3code) - Open source AI development environment
- [Codex](https://openai.com/codex) - OpenAI's code generation model and tools
- [Cursor](https://cursor.sh) - AI-powered code editor
- [Claude Code](https://claude.ai) - Anthropic's coding assistant
## 🗣️ Paraphrase
A developer completely overhauled their AI coding workflow over five months while building a new full-stack framework called Lakebed. The major shift involved abandoning Cursor's plan mode and Claude models in favor of GPT-4.5 with simpler tooling approaches.
The new workflow emphasizes remote development capabilities, allowing work to continue seamlessly when disconnecting from the primary machine. Rather than complex configuration files and parallel work streams, the approach focuses on single-threaded development with minimal setup and natural conversation-based steering of AI models.
The key insight is treating AI collaboration as a conversation requiring careful reading of outputs and thoughtful responses, rather than trying to engineer perfect prompts upfront. Simple voice-to-text planning and screenshot-based feedback proved more effective than elaborate technical specifications.
## 📰 Newsletter
**This Week in AI Development: The Great Workflow Shake-Up**
A fascinating deep-dive into how quickly AI development practices are evolving! One developer completely rewrote their approach in just 5 months, moving from Cursor's plan mode to a GPT-4.5 + remote development setup that's worth studying.
Key takeaways for WordPress developers: The shift toward **conversation over configuration** mirrors what we see in modern WordPress development - simple, focused tools often outperform complex setups. The emphasis on **remote development** could be game-changing for agencies managing multiple client environments.
The creator's "Lakebed" project (a new full-stack framework) demonstrates how AI can tackle ambitious technical projects when properly steered. Their agent.md philosophy - writing it like "a letter to the agent" explaining project goals rather than technical constraints - offers a fresh approach to AI collaboration.
**Tools mentioned**: T3 Code (open source), Codex app, Tailscale for secure remote access. The remote coding setup enabling seamless work across devices feels particularly relevant as our industry becomes increasingly distributed.
Worth watching if you're exploring AI-assisted development workflows or curious about the rapid evolution of developer tooling.
## ✅ Action Items
#### DevRel
- Research T3 Code integration for developer workshops on AI-assisted coding
- Create demo showcasing remote development workflows for distributed teams
- Document voice-to-text prompting techniques for technical planning sessions
#### WordPress
- **Core:** Evaluate AI-assisted development patterns for core contribution workflows
- **Community:** Develop guidelines for AI tools in WordCamp workshop formats
- **Training:** Create curriculum on modern AI development practices for WordPress teams
- **CLI:** Explore AI integration patterns similar to Codex CLI approaches
#### Community
- **Maintainers:** Document best practices for AI-assisted project planning and execution
- **Contributors:** Establish guidelines for AI tool use in collaborative development
- **Ecosystem:** Evaluate remote development infrastructure for distributed WordPress teams
## 🧪 Hypotheses
1. **Conversation-driven AI collaboration produces better results than elaborate prompt engineering**
2. **Remote development infrastructure will become standard for AI-assisted coding workflows**
3. **Single-threaded AI development tasks are more productive than parallel approaches**
4. **Voice-to-text planning creates superior initial context compared to written prompts**
5. **Minimal agent configuration files outperform comprehensive technical specifications**
### 📊 Methodology
**Study Design**: Comparative workflow analysis tracking developer productivity metrics across different AI coding approaches
**Data Sources**:
- Development velocity measurements (features per day, bugs per session)
- Context switching frequency and task completion rates
- Remote vs local development session quality metrics
- Prompt complexity vs output satisfaction correlation
**Key Variables**: Model choice (GPT-4.5 vs Claude), environment (Cursor vs T3 Code), work patterns (single vs multi-threaded), configuration complexity
### ❓ Questions
- How does voice-to-text prompting compare quantitatively to written prompts for development planning?
- What are the security and compliance implications of remote AI development workflows?
- How do different AI models handle large codebase context management?
- What factors determine when to create new threads vs continue existing conversations?
- How does screenshot-based feedback compare to written descriptions for UI changes?
### 🔮 Future Research
- Longitudinal studies comparing developer productivity across AI workflow approaches
- Analysis of AI model performance degradation with context length in coding tasks
- Security framework development for remote AI-assisted development environments
- Comparative analysis of different agent configuration philosophies and outcomes
- Investigation of voice interface patterns for technical communication with AI systems
### 💡 Implications
**Practice**: Development teams may need to fundamentally rethink project management approaches for AI-assisted workflows. The single-threaded, conversation-driven model contradicts many current parallel development practices.
**Tooling**: Infrastructure providers should prioritize remote development capabilities and seamless device switching. Traditional IDE-centric approaches may become less relevant.
**Education**: Developer training programs need to incorporate AI collaboration techniques, emphasizing communication skills alongside technical knowledge.
**Security**: Remote AI development introduces new attack vectors requiring updated security frameworks and compliance approaches.
### 🧠 Learning Styles Assessment
**Reading Level**: Flesch-Kincaid Grade Level ~12-14. Technical content with domain-specific terminology but conversational tone makes it accessible to experienced developers.
#### 👀 Visual learners
**Present**: Screenshots of development environments, workflow diagrams, before/after interface comparisons
**Improve**: Architecture diagrams of remote development setup, visual comparison charts of different AI tools, annotated screenshots showing specific workflow steps
#### 🔊 Aural learners
**Present**: Original video format with detailed narration
**Improve**: Podcast version focusing on key insights, audio summaries of each major workflow change
#### 🗣️ Verbal learners
**Present**: Clear definitions of technical terms, analogies for complex concepts
**Improve**: Glossary of AI development terminology, plain-language explanations of technical architecture
#### 🧩 Kinesthetic learners
**Present**: Step-by-step setup instructions, hands-on tool configuration
**Improve**: Interactive tutorials, downloadable configuration files, guided practice exercises
#### 🧠 Logical learners
**Present**: Decision trees for tool selection, systematic comparison of approaches
**Improve**: Flowcharts for workflow decisions, quantified performance metrics, cause-and-effect analysis
#### 🧑🤝🧑 Social learners
**Present**: Community discussion around tool choices and workflows
**Improve**: Team collaboration guidelines, peer review processes for AI-assisted code
#### 🧘♀️ Solitary learners
**Present**: Self-guided experimentation with different tools and approaches
**Improve**: Personal reflection prompts on workflow effectiveness, individual assessment rubrics
#### 🌿 Naturalistic learners
**Present**: Evolution metaphors for workflow development
**Improve**: Ecosystem thinking about tool integration, sustainable development practices
### 📊 Accessibility Assessment
#### Screen reader semantics
**Present**: Well-structured headings and clear navigation
**Improve**: ARIA labels for technical diagrams, alt text for screenshots describing specific UI elements
#### Keyboard navigation
**Present**: Standard web navigation patterns
**Improve**: Skip links to major sections, keyboard shortcuts for video controls
#### Contrast and readability
**Present**: Good text contrast ratios
**Improve**: High contrast mode for code examples, larger font options for technical content
#### Text resizing and reflow
**Present**: Responsive design supports zoom to 200%
**Improve**: Improved line spacing for code blocks, better mobile text flow
#### Media and captions
**Present**: Video content with clear audio
**Improve**: Full transcript with timestamps, closed captions for technical terms
### 🌍 DEIB Assessment
#### Cultural breadth
**Present**: English-focused content with some universal technical concepts
**Improve**: Multi-language tool examples, international developer community perspectives
#### Inclusive language
**Present**: Generally neutral technical language
**Improve**: Replace "guys" with inclusive alternatives, define technical jargon consistently
#### Neurodiversity support
**Present**: Clear section breaks and structured information
**Improve**: Content warnings for rapid context switching discussions, consistent formatting patterns
#### Representation
**Present**: Focus on technical content rather than personal identity
**Improve**: Acknowledge different learning and working styles, include accessibility considerations in tool recommendations
#### Psychological safety
**Present**: Respectful tone toward different technical approaches
**Improve**: Explicitly acknowledge that different workflows suit different people, avoid dismissive language
### ✅ Logical Validity
#### Strong reasoning
**Deductive strength**: The argument that simpler AI configurations outperform complex ones is well-supported by specific examples and measurable outcomes during the Lakebed project development.
**Inductive evidence**: Multiple concrete examples (environment variables, user limits, HTML planning) demonstrate consistent patterns of success with the new approach.
**Causal reasoning**: Clear before/after comparison shows workflow changes coincided with improved productivity, though correlation vs causation could be better established.
#### Fallacies
**Appeal to personal experience**: Heavy reliance on individual experience without acknowledging that different developers may have different optimal workflows. **Fix**: Include caveats about individual variation and context-dependency.
**Hasty generalization**: Concluding that SSH terminal coding is universally inferior based on limited use cases. **Fix**: Acknowledge legitimate use cases while highlighting limitations.
**False dichotomy**: Presenting plan mode vs conversation as binary choice rather than potentially complementary approaches. **Fix**: Explore hybrid approaches and context-appropriate tool selection.
#### Tone of voice
**Confident and opinionated** with casual, conversational style. Uses strong language ("mental illness," "insulting") to emphasize points but maintains technical credibility through detailed examples and reasoning.
### 🛡️ DARVO Analysis
No clear DARVO patterns detected. The content critiques tools and approaches but doesn't involve accusations or victim/offender dynamics. The author takes responsibility for their previous recommendations and acknowledges changing perspectives based on new experience.
## 🎯 Conclusion
This workflow evolution represents a significant shift toward simpler, conversation-driven AI development practices. The emphasis on remote capabilities, single-threaded focus, and minimal configuration suggests the field is maturing beyond early experimentation toward practical, sustainable approaches.
**Key practical takeaways**: Read AI outputs thoroughly, steer through conversation rather than complex prompts, prioritize remote development infrastructure, and treat each task as a focused conversation rather than part of a complex multi-threaded workflow.
The rapid pace of change (complete workflow overhaul in 5 months) highlights how quickly AI development practices are evolving and the importance of remaining adaptable to new approaches.
## 🔄 Refactor Appendix — How I code with AI changed a lot
### 📚 Literature Atoms (H3)
#### GPT-4.5 outperforms Claude for sustained development work
> [!quote] ID:rw_transcript_001
> "I really struggled to use anything other than GPT 5.5... I've pretty much entirely stopped using clawed models"
**Context & use**:
- Model comparison after 5 months of intensive development
- Sustained performance over multiple coding sessions more important than peak capabilities
**Tags**: #xref #ai-models
**Highlight**: [view highlight](readwise://transcript_timestamp_001)
<!-- atomicity: 1 idea ✓ -->
#### Conversation-driven steering outperforms complex prompt engineering
> [!quote] ID:rw_transcript_002
> "The most important thing when you're building with AI is that the AI understands what you want and how you build... just read the outputs and steer the model in the direction you want to go"
**Context & use**:
- Core philosophy shift from elaborate configuration to natural interaction
- Emphasis on reading AI text outputs, not just generated code
**Tags**: #xref #ai-collaboration
**Highlight**: [view highlight](readwise://transcript_timestamp_002)
<!-- atomicity: 1 idea ✓ -->
#### Remote AI development enables seamless device switching
> [!quote] ID:rw_transcript_003
> "I want to be able to run the thing, close my laptop or go offline or code from an Uber and then have it updated reliably"
**Context & use**:
- Remote development infrastructure as core productivity enabler
- Addresses fundamental limitation of local-only AI coding workflows
**Tags**: #xref #remote-development
**Highlight**: [view highlight](readwise://transcript_timestamp_003)
<!-- atomicity: 1 idea ✓ -->
### 🧠 Permanent Candidates (H4)
#### AI collaboration requires reading outputs before examining code
**Claim**: Developers must prioritize understanding AI reasoning and explanations over immediately reviewing generated code to effectively steer AI development conversations.
**Evidence pointer**: [[#Conversation-driven steering outperforms complex prompt engineering]]
**Scope & conditions**:
- Holds when: Doing complex, multi-step development tasks requiring iteration
- Breaks when: Simple, single-shot code generation where output quality is immediately apparent
**Links**:
- [[AI development workflow]] — because extends beyond simple code generation
- [[Context management]] — because contrasts with prompt-focused approaches
- [[Developer-AI collaboration]] — because applies-in sustained project development
<!-- atomicity: 1 idea ✓ -->
### ✏️ Fleeting Queue
- Voice-to-text planning creates better initial context than written prompts — investigate Whisper Flow integration patterns #fleeting #inbox
- T3 Code remote hosting setup via Tailscale for team development environments — https://github.com/t3-oss/t3code #fleeting #inbox
- Agent.md as "letter to the agent" vs technical specification approach — compare effectiveness #fleeting #inbox
### → Things 3 (max 2 actions)
- [ ] Research T3 Code setup and remote development capabilities for team workflows #things
- [ ] Document voice-to-text prompting techniques for AI development planning sessions #things
<!-- deep-read:sections end -->
## Highlights
- applies single developer's experience to all AI-assisted development
- Note: 🔍 Fallacy: Hasty Generalization #fallacy/hasty-generalization