my-ai-toolsDocumentation index

My AI Coding Tools Learning Notes

Personal notes from experimenting with AI coding tools, workflows, and frameworks.

Keybinder

A modern, intuitive macOS app for managing skhd keyboard shortcuts.

Features

  • Auto-Detection - Finds skhd config from standard locations
  • Visual Editor - Clean interface for editing keyboard shortcuts
  • Safety Controls - Confirmation for destructive commands
  • Real-time Log Viewer - Live streaming of service logs
  • System Theme Integration - Automatic light/dark mode

Development Approach

Used spec-kit for specification-driven development:

  1. Initial feature planning with structured specs
  2. Implementation tasks generated from specifications
  3. Pragmatic deviation during actual development (specs as reference, not strict rules)

SealCode (VS Code Extension)

Smart Code Review with AI-Powered Insights

Features

  • AI-Powered Review - Send comments to Claude, Copilot, OpenCode, or Amp
  • Prompt Templates - Built-in templates for review, security, refactor workflows
  • Categorized Comments - Bug, Question, Suggestion, Nitpick, Note
  • Rich Visual Feedback - Inline decorations, gutter icons, line backgrounds
  • Export Options - Export reviews to Markdown or HTML

Development Approach

Built autonomously using Ralph (AI agent system) with Amp as the coding assistant:

  • PRD-driven development with task decomposition
  • Autonomous implementation of extension features
  • Integrated testing and quality validation

Ralph

Autonomous AI agent loop for PRD-driven development.

Minimal Ralph implementation: PRD → task decomposition → autonomous execution → loop until complete. Used in SealCode development. Will deprecate once upstream PRs land.

AI Launcher

Fast launcher for switching between AI coding assistants.

Fuzzy search interface for Claude Code, OpenCode, Amp, etc. Quick switching without managing multiple terminals.

Tiny Coding Agent

Minimal coding agent focused on simplicity.

Lightweight agent with minimal dependencies. Response to heavy frameworks that waste tokens on orchestration overhead.

HermesHub

Web app for deploying and managing a self-hosted Hermes AI Agent on any VPS — no terminal required.

  • Repository: jellydn/hermes-hub
  • Website: hermes-hub.itman.fyi
  • Built with: GPT-5.5 (planning) + Grok CLI Composer 2.5 (implementation)
  • Tech Stack: TanStack Start, Tailwind CSS v4, shadcn/ui, Hono, PostgreSQL, Drizzle ORM, Better Auth, Vitest

Features

  • Guided VPS setup — Step-by-step server connection wizard for non-technical users
  • One-click Hermes deployment — Install Docker, Compose, and Hermes from the dashboard with live SSE progress
  • AI provider configuration — OpenAI, Anthropic, OpenRouter, Ollama, and custom endpoints without editing env files
  • Telegram onboarding — Connect a bot, verify the token, and approve pairing codes from one screen
  • Agent persona editor — Define how Hermes speaks via SOUL.md, then deploy to a chosen VPS
  • MCP server manager — Add stdio or HTTP MCP servers (with presets) and deploy to Hermes config.yaml
  • Built-in Hermes Web UI — Deploy and open the Hermes browser interface from the server detail page

Development Approach

Split planning and implementation across two models:

  1. GPT-5.5 for planning — Architecture, API design, feature specs, and execution plans before writing code
  2. Grok CLI Composer 2.5 for implementation — Fast iteration on TanStack Start routes, Hono API handlers, SSH deploy pipelines, and Vitest coverage

The planner produced bite-sized tasks with file paths and verification steps; Composer 2.5 executed them in focused sessions. Review and docs passes stayed in the same Grok CLI workflow.

Key Takeaways

Tool Combination Best For
Claude + spec-kit Greenfield projects requiring structured planning
GPT-5.5 + Grok CLI Composer 2.5 Plan-heavy web apps with fast, tool-aware implementation
Amp + Ralph Autonomous development with PRD-to-implementation pipelines
AI CLI Switcher Developers working with multiple AI tools
Tiny Agent Cost-conscious development with minimal overhead

Focused, single-purpose solutions > heavy, all-in-one frameworks.

Tools I Tried But Didn't Keep

Not every tool fits every workflow. Here are tools I evaluated but chose not to adopt:

Task Master (claude-task-master)

Repository: eyaltoledano/claude-task-master

An AI-powered task management system designed for Cursor, Windsurf, and other editors.

Why I didn't keep it:

  • Complexity - Requires multiple API keys (Anthropic, OpenAI, Perplexity, etc.)
  • Heavy setup - MCP configuration, environment variables, PRD structure
  • Overkill - 36 tools available, ~21,000 tokens context usage in full mode
  • Too structured - Strict PRD-to-task workflow doesn't fit iterative development

SuperClaude

Website: superclaude.netlify.app

A meta-programming configuration framework for Claude Code.

Why I didn't keep it:

  • Context hungry - Uses too much context for the framework overhead
  • Mental overhead - Learning the framework's abstractions vs just coding
  • Over-engineered - Adds complexity without proportional benefit

Oh My OpenCode

Repository: code-yeongyu/oh-my-opencode

A feature-rich plugin for OpenCode with multi-model orchestration, background agents, and "Sisyphus" workflow.

Why I didn't keep it:

  • Too opinionated - Enforces specific workflows (ultrawork, ralph-loop, Sisyphus mode)
  • Multi-model complexity - Requires multiple API keys (Claude, GPT, Gemini, Grok)
  • Magic keywords - ultrawork, ulw, ultrathink - adds cognitive overhead
  • Feature bloat - LSP tools, AST-grep, background agents, keyword detectors
  • Not my style - Aggressive automation doesn't fit my iterative workflow

Addy Osmani — Accountable Engineering with AI

Video: The Future of Software Engineering with AI

Addy Osmani (Chrome, DevTools, Core Web Vitals, AI developer experience) argues AI will not eliminate software engineering — it changes what valuable engineering looks like.

Key Takeaways

  • AI expands the software market — when building got easier historically, more people built more software; expect expansion, not just job reduction.
  • Accountability shifts — AI generates code; humans remain responsible for correctness, security, maintainability, and fitness for users.
  • Avoid cognitive surrender — you do not need every line inspected, but you must understand major architectural, security, and business decisions.
  • Mutual amplification — agents record decisions and lessons; developers review and internalize them. The goal is a better feedback loop, not only faster output.
  • Management is technical again — AI lowers the barrier to building; senior leaders are coding weekend projects again.
  • Memory debugging lags — profiling and tracing improved; memory management tooling has not advanced as much.
  • Broaden beyond coding — product thinking, UX, evangelism, go-to-market, and customer understanding increasingly matter.

Practical Workflow (Captured as a Skill)

The accountable-engineering skill turns these lessons into one authoritative eight-step workflow with explicit completion criteria from definition through operational ownership.

Senior AI Engineer Profile

Own complete outcomes, not just AI features:

Dimension What it means
Technical depth RAG, agents, evaluation, deployment, observability
AI leverage Use coding agents; verify their work
Product sense Know who benefits and what success means
Operational ownership Reliability, cost, security, monitoring
Communication Explain trade-offs to any audience

The valuable engineer is an accountable system builder, not merely a fast coder.

Worth Learning From

These tools aren't for me daily, but are valuable references for understanding AI-assisted workflows:

Superpowers

Repository: obra/superpowers

A complete software development workflow for coding agents with composable "skills".

Key concepts worth studying:

  • Brainstorming phase - Agent asks questions before coding, refines specs iteratively
  • Writing plans - Bite-sized tasks (2-5 min each) with exact file paths and verification steps
  • Subagent-driven development - Fresh subagent per task with two-stage review
  • TDD enforcement - RED-GREEN-REFACTOR cycle, deletes code written before tests
  • Git worktrees - Isolated workspaces on new branches

Why it's a good reference:

  • Well-structured skill composition patterns
  • Philosophy: systematic over ad-hoc, evidence over claims
  • Shows how to build layered agent workflows

My Philosophy: Keep It Simple

"Add only if needed"

Instead of adopting heavy frameworks, I prefer:

Approach Why
Minimal tooling Less context usage, more tokens for actual work
AGENTS.md Simple, portable project guidance
Native AI features Use built-in Claude/Amp capabilities first
Add incrementally Only add tools when there's clear friction

The best tool is the one you don't have to think about.