Personal learning repository tracking AI-assisted development across the full spectrum—from autocomplete to autonomous agents.
AI-assisted coding has moved well beyond autocomplete. The tooling landscape now spans IDE completions, chat-based problem solving, boilerplating, code generation, and spec-driven autonomous development. Tools like Kiro, Kilo Code, Cline, and others are pushing toward agents that can plan, execute, and iterate with decreasing human intervention.
That said, the models—even with the impressive coding skills of the newest generation—still require human review, domain expertise, and polish. Production code needs coders who understand what they're building. Even the most capable autonomous agents work best with a human in the loop for architecture decisions, edge cases, and quality gates.
The skeptics who dismissed AI coding entirely are now using autocomplete; the early adopters are experimenting with spec-driven workflows. The whole field is moving, and it's moving fast.
This repository is my personal focus for 2026: building fluency across these tools as they evolve. It consolidates projects that would otherwise sprawl into separate repos—learning vehicles, experiments, and practical builds that don't yet warrant their own home. When something graduates to "project worthy," it moves out. Everything else lives here.
Looking back at this repo in December 2026 should be an interesting view of how the year unfolded.
year-of-code-2026/
├── ai-agentic-coding/ # Spec-driven development and CLI agent experiments
├── ai-assisted-research/ # NSB methodology for AI research agents
├── ai-taskmaster/ # AI competitive game show concept
├── assets/ # Repository-level images and banners
├── aws-space-weather-dashboard/ # Serverless NOAA data application
├── docs/ # Documentation standards and templates
├── ml-gymnasium-explainable-ai/ # Explainable RL with Gymnasium environments
├── ml-orbweaver-roguelite/ # Monte Carlo roguelite mechanics validation
├── work-logs/ # Milestone worklogs and weekly rollups
├── CONTRIBUTING.md
├── LICENSE
└── README.md| Project | Domain | Status | Progress | Current Phase |
|---|---|---|---|---|
| aws-space-weather-dashboard | AWS/Serverless | 🔄 Active | Spec 01/11 | Identity & Auth |
| ml-orbweaver-roguelite | ML/Game Design | ✅ Sweep Ready | Engine Complete | Production Sweeps |
| ml-gymnasium-explainable-ai | ML/Education | 🔄 Active | env01 Complete | YouTube Production |
| ai-assisted-research | Methodology | 🔄 Active | v0.4 Complete | Example Prompts |
| ai-agentic-coding | Tooling | 📁 Ongoing | — | Collection |
| ai-taskmaster | Creative | ⏸️ Blocked | Concept Complete | TTS Selection |
Legend: 🔄 Active Development | ✅ Milestone Complete | 📁 Ongoing Collection | ⏸️ Blocked
A full-stack serverless application that ingests NOAA GOES satellite data, stores it in a dual-database architecture (DynamoDB for fast access, RDS PostgreSQL for historical queries), and serves it through an authenticated API to a React dashboard. The real purpose is learning production AWS patterns by building something real—the stack mirrors a customer engagement I'm preparing for, with the same services, architecture decisions, and operational concerns.
The project follows a 12-spec sequence from foundation through observability, built entirely through Kiro's spec-driven autonomous workflow with the constraint of never manually touching the AWS console. Spec 00 (Foundation) established CloudTrail, IAM baseline, and cross-stack sharing patterns that carry through the remaining specs.
Stack: CDK TypeScript Lambda API Gateway Cognito DynamoDB RDS S3 EventBridge React Material UI
AI Tooling: Claude Opus 4.5, Kiro IDE, AWS MCP
Status: 🔄 Active (Spec 01/11 — Identity & Auth)
A headless Python simulation validating roguelite game mechanics through Monte Carlo analysis before committing to Godot engine implementation. The core thesis: every bullet-hell roguelite ends the same way—you're cornered, you plant your feet, you fight until you win or die. This simulation skips to that end state and asks whether strategic decisions alone (via a "hate draft" mechanic where player and room draft from a shared modifier pool) can carry engaging gameplay.
The simulation runs 10K+ games with deterministic seeding, tracking win rates, wave survival curves, and draft decision patterns. A complete parameter sweep engine using Sobol sequence sampling and Wilson CI adaptive stopping rules enables exploration of the viable parameter space. The current v5_lowturret baseline achieves 53.5% player win rate with a smooth difficulty curve peaking at wave 10.
Stack: Python 3.11+ dataclasses CSV scipy numpy
AI Tooling: Kiro IDE (~$2.78 total for simulation + sweep engine)
Status: ✅ Sweep Ready (v5_lowturret baseline, sweep engine complete)
Companion code for a YouTube series on explainable reinforcement learning, focused on understanding why agents make decisions rather than just that they work. The approach: train agents on Gymnasium environments, then crack them open with Q-value overlays, policy inspection, and decision visualizations. Each environment gets its own mini-series on ClusterForge ML.
The xrl library provides reusable inspection and visualization tools across environments—DQN Q-value extraction, advantage calculation, HUD overlays, and Q-bar renderers. CartPole (env01) is complete with training scripts, explained video recording, and analysis outputs. The pattern extends to Lunar Lander, MiniGrid, and eventually continuous control environments.
Stack: Python 3.10+ Gymnasium Stable-Baselines3 PyTorch OpenCV
AI Tooling: Claude Opus 4.5
Status: 🔄 Active (env01-cartpole complete, YouTube production pending)
Methodology for getting better outputs from AI research agents, developed through 1000+ Gemini Deep Research sessions. The core framework is Negative Space Bounding (NSB)—a three-layer constraint architecture that collapses the search space before agentic iteration begins rather than filtering sprawl afterward.
The layered approach uses Anchors (immutable physical/economic context), Walls (hard exclusions that prevent search vectors from being generated), and Gates (conditional logic that evaluates utility during synthesis). The v0.4 methodology synthesizes earlier walls-only and gates-only approaches into an integrated architecture that transfers to any agentic research system.
Stack: Gemini Deep Research, prompt engineering
AI Tooling: Gemini 2.5/3.0 Deep Research
Status: 🔄 Active (v0.4 methodology complete, example prompts pending)
Catch-all for experiments, observations, and one-offs from the spec-driven and autonomous end of AI-assisted development. The territory between "AI autocomplete" and "fully autonomous agent" is wide and weird—this folder tracks working in it. Content includes spec-driven development patterns with Kiro, CLI agent sessions with Claude Code and Gemini CLI, workflow experiments, and methodology notes on prompt engineering and human-in-the-loop patterns.
The larger projects in this repository (aws-space-weather-dashboard, ml-orbweaver-roguelite) use agentic tooling extensively; this folder captures meta-level learnings from that work. CodeRabbit was selected as the primary AI code review integration after evaluating several options.
Stack: Various
AI Tooling: Kiro, Claude Code, Codex CLI, Gemini CLI, Kilo Code, Cline
Status: 📁 Ongoing collection
A competitive game show format where frontier AI models compete in ambiguous, bespoke tasks—judged by an imperious Taskmaster (Claude Opus 4.5). The innovation is a dual-voice system: each AI contestant has both a "table voice" (what they say to other contestants, strategic and performative) and an "audience aside" (fourth-wall address revealing true strategy). The comedy emerges from the gap between public performance and private commentary.
Think Taskmaster meets QI—four AI models (Claude Sonnet, GPT-5, Gemini 3 Pro, rotating fourth) competing in tasks like "make the Taskmaster laugh in exactly 17 words" or "convince the judges you're the only real AI." Production approach uses visual novel layout with TTS voices and generated scene images. Concept doc and character avatars are complete; blocked on TTS voice synthesis selection.
Stack: Visual novel layout, TTS (pending), image generation
AI Tooling: Multiple frontier models as contestants, Claude Opus 4.5 as Taskmaster
Status: ⏸️ Concept complete, blocked on TTS selection
Naming: Projects are prefixed by domain—aws-, gcp-, azure-, ml-, ai-, etc.
Structure: Each project maintains its own README with goals, approach, and status.
Memory Banks: Projects use .kilocode/rules/memory-bank/ for persistent context—brief, architecture, product, tech, context, and tasks files that AI agents load at session start.
Promotion: Projects graduate to their own repository when they become "project worthy"—a judgment call based on scope, external collaboration, or standalone value. Size alone doesn't determine this.
Sensitive Data: This is a public repo. Credentials, secrets, and sensitive configuration live in a separate private repository with manual coordination.
This repository benefits from open source programs that provide free or discounted tooling to qualifying public repositories. These aren't direct sponsorships—they're programs that support the open source ecosystem, often auto-approving based on repository metrics. The companies may not know this specific repo exists.
| Program | Provides | Use Case |
|---|---|---|
| CodeRabbit | AI code review (Pro tier) | PR review with codebase context, CLI integration with Claude Code and Gemini CLI |
| Atlassian | Jira, Confluence, Bitbucket Pipelines (Standard tier) | Project tracking, documentation, CI/CD |
| Program | Provides | Planned Use |
|---|---|---|
| Snyk | Security scanning | Dependency vulnerability detection |
| SonarCloud | Code quality analysis | Static analysis, code smell detection |
| Sentry | Error tracking | Runtime error monitoring |
| Datadog | Observability | Metrics, logs, APM |
MIT © 2025 VintageDon
Some projects in this repository accept contributions. See CONTRIBUTING.md for which projects are open and how to submit a PR.
AI-assisted contributions are welcome—this repo is about exploring those tools, after all. But your code must be good, and you must understand what you're submitting.
GitHub: @vintagedon