You've Been Using AI the Hard Way (Use This Instead)

About this video

NetworkChuck argues that terminal versions of AI tools beat the browser apps, which scatter work across chats that lose context. He demos Google’s Gemini CLI first (generous free tier, Gemini 2.5 Pro): installed with one command, it answers questions, searches the web, and — the key difference — reads and writes files directly, creating coffee-research documents on request. Its /init command generates a GEMINI.md context file by analyzing the project folder, so new sessions reload the project’s state instead of starting over. His main tool is Claude Code: he shows its CLAUDE.md equivalent, /context token visibility, custom agents created via /agents (specialists like a ‘home lab guru’ and a deliberately harsh ‘brutal critic’) that run as separate instances with fresh context windows so the main conversation stays lean, parallel agents, resuming sessions, a dangerously-skip-permissions mode, plan mode, and output styles that swap its persona/system prompt (he writes this video’s script with one). He runs Claude Code, Gemini CLI, and OpenAI’s Codex side by side in one folder, keeping GEMINI.md, CLAUDE.md, and AGENTS.md synced, uses a ‘session closer’ agent that summarizes the day, updates all context files, and commits the project to GitHub like code, and finishes with OpenCode, an open-source alternative that can run local models and can use a Claude Pro login. A sponsor segment for Twingate makes the security point that giving terminal AI file/network access makes traditional VPNs risky.

Key takeaways

• The core shift is owning your context: project state lives in local files (GEMINI.md, CLAUDE.md, AGENTS.md) on your own disk, portable between tools and vendors, instead of being trapped in browser chat histories.
• Terminal AIs act on the machine — reading and writing files, running bash/Python — so deliverables are created in place rather than copy-pasted out of a chat window.
• Sub-agents protect the main context window: in his demo a research agent burned about 40,000 tokens on its own fresh window while the main conversation only grew by roughly 4,000 tokens.
• Running rival tools together works because they share the folder: he has Claude draft, Gemini draft from another angle, and Codex critique, with a session-closer agent syncing all three context files and committing to Git at day’s end.
• Access and cost notes from the creator: Gemini CLI is his free starting recommendation; Claude Code is paid but, he says, a Claude Pro subscription (starting around $20/month by his account) logs in without API keys; OpenCode is the open-source option that can also run local models.

Episode details: published October 28, 2025 · 33:44 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are NetworkChuck’s own claims, shared for education — they are not typical or guaranteed.

← Previous: This AI Supercomputer can fit on your desk… · Next: DEFCON is Not What I Expected… →

All NetworkChuck videos · NetworkChuck profile · NetworkChuck on YouTube