Me, Myself, and AI: 5 Days, 2 Models, 200M Tokens

I just spent 5 days building with AI coding assistants. The first time going from an empty repo to a working platform with my AI friends. The setup: two models, two coding tools, and a workflow that kept them both honest.

The Workflow

Write specs first → 14 detailed markdown docs (API contracts, database schemas, workflows)

Build with multiple models:

• Cursor + GPT-5 Codex → Initial scaffolding and feature builds
• Cursor + Claude 4.5 → Refactoring and iteration (noticeably faster than Claude 4)
• Augment Code + Claude 4.5 → Code review and debugging

Review loop:

• One model builds a feature → Switch tools → Other model reviews it
• Hit a bug → Get fix from Model A → Ask Model B to critique the approach
• Models disagree → Forces you to actually understand the tradeoffs

---

Patterns That Worked

Specs as shared context:

• Both models reference the same requirements
• Prevents drift and hallucinations
• Serves as documentation

‍Tool strengths:

• Augment Code → Deep debugging and the best at following my specs
• Claude 4.5 → Faster (MUCH faster) iteration, good diagnosis and tracking down of issues
• GPT-5 Codex → Complex implementations

It could have just been because I was running the full stack locally in docker…but Augment Code sometimes would slow down dramatically, something to look at: the memory footprint of VScode w/ Augment vs Cursor. Cursor seemed to keep up and move quickly even as the repo and context got bigger.

Having models disagree catches more bugs than either alone.

---

By the Numbers:

• 5 days → concept to working system
• 200M tokens → back-and-forth iteration
• 14 spec docs → written before major features
• 138 commits → small, iterative changes
• 2 coding tools → Cursor + Augment Code
  

---

Bottom line: The tools are powerful, but you need structure. Specs keep models honest. Multiple models catch more issues. Claude 4.5’s speed is a huge bump from 4. Codex does keep up pretty well though.

The engineering still matters - you need to architect properly and know when AI is wrong. But with the right workflow, you can move incredibly fast.

---

The (Maybe Over-Engineered) Stack

Backend:

• FastAPI → REST API and async processing
• CrewAI → Multi-agent orchestration (5 agents)
• PostgreSQL → Structured data
• Neo4j → Graph relationships
• Redis → Caching and task queues
• FAISS → Vector Store

Frontend:

• React + TypeScript
• Tailwind CSS 

Infra:

• Docker + Docker Compose → Deployment
• Celery → Background task processing

Could I have simplified this? Probably. The models handled it fairly well. The specs helped A LOT to keep everything aligned with what I designed.I’m iterating on some other ideas for managing these agents as I clean it up and get it ready for production. Overall…me, myself, and AI did a great job!