Build Interchangeable Directories as Your Core OS

Treat coding agents like Claude Code, Codex, Hermes Agent, and OpenClaw as swappable 'harnesses' that plug into persistent project directories (e.g., 'herk-2' with .claude.md files, scripts, skills). This makes your workflow tool-agnostic: if Claude Code shuts down, swap in Codex or another without rebuilding. Directories outlive tools, ensuring new agents integrate seamlessly. Extract features from specialized tools (e.g., NotebookLM) into your custom ecosystem for customization and cost savings.

S-Tier Daily Drivers (live in these 100% of the time):

  • Claude Code: Primary OS for all work; handles coding, agents, automations.
  • VS Code: IDE host for Claude Code (via extension/terminal); pairs with Cursor or Windsurf alternatives.
  • Glido: Fastest private speech-to-text (replaced Whisper); agentic features incoming, Windows support soon (free month via link).

A-Tier Weekly Use (complements S-tier):

  • Codex: Pairs with Claude Code to cover weaknesses.
  • Claude Chat: Quick chats when not in Claude Code.
  • Hermes Agent: On-demand via Telegram for mobile/general knowledge; instant crons, easy setup.
  • Perplexity: Agent research.
  • Grok (in X): Twitter thread insights/search.

B-Tier Specialists (task-specific):

  • Apify: Scraping/actors for agents.
  • GPT Image 2: Creative image gen.
  • NanoBanana 2: Photoshop-like edits.
  • Key.AI: Image/video model router for agents.
  • OpenRouter: Model routing.
  • HeyGen: Avatars (e.g., course videos).
  • 11 Labs: Voice cloning/agents.
  • Claude Design: Team landing pages with shared design system.

C-Tier Experimenting:

  • Gemini/Anti-Gravity: Rare use.
  • Ollama: Open-source model testing/cloud.
  • Manus: Occasional tests (great for AI newbies as S-tier).

Graduated (extracted/replaced, not trash):

  • ChatGPT, OpenClaw (Hermes replaced), Cursor, NotebookLM, NotebookAI, Whisper (Glido replaced), Poppy AI (replicate in Claude Code).

Non-AI supports: Hostinger VPS (NATEHERK 10% off annual), ClickUp (PM), Fireflies (meetings).

This lean core (Claude Code + Glido) handles full days; specialists slot in per task.

Decision Framework: Test Only Real Pain Solvers

When a new tool/feature drops (e.g., YouTube video):

  1. Does it solve a current pain point? (Usually no—save link.)
  2. If yes, test in real low-risk scenario (not mock data) for 1 week.
  3. Evaluate: Solves pain enough for main stack? Keep or discard.

Revisit saved links only at roadblocks: if a tool clears it, learn then. Stay on north star path (e.g., business mission)—new releases distract unless aligned. Knowing 'what' (10-min video) beats 'how' (full build) for most.

Productivity Rules to Maximize Needle-Moving Output

  • 20% Dip Rule: Switches cause ~20% efficiency drop; justify only if post-dip exceeds prior baseline (blue line > green; avoid red flat recovery).
  • Needle per Hour > Hours Worked: Prioritize north star goal daily (e.g., 'Achieve X by EOD'); 4 focused hours > 12 scattered (threads/posts/videos). Experiment post-goal.
  • Task-Level Tool Picking: Break processes into mini-tasks (e.g., YouTube: Perplexity research → Claude Code structure → Claude Chat script → GPT Image 2 thumbnail → NanoBanana edits → Premiere). Pick best tool per step; mix AI/non-AI.

Bezos principle: Focus on unchanging (directories, processes) over trends. Tool dependence kills flow—if Claude down, swap seamlessly.