RAW Agentic Coding: ZERO to Agent SKILL

Quick Overview

The creator builds a custom, reusable agent skill named "Fork Terminal Skill" from scratch using a structured Markdown format that defines purpose, variables, instructions, workflow, and cookbooks for specific LLMs (Claude, Codex, Gemini) to enable agents to fork new terminal windows for specific tasks like running commands or summarizing codebases, demonstrating its capability by successfully executing a command and then demonstrating context-aware summarization for different models.

Key Points: The core concept demonstrated is building an agent skill from scratch using a structured Markdown file, where the skill is designed to fork new terminal sessions for agents. The skill documentation outlines specific instructions for Claude, Codex CLI, and Gemini CLI agents, including model preferences (e.g., Claude: Haiku/Sonnet, Codex: gpt-3.5-codex-mini, Gemini: gemini-2.5-flash) and required flags like or . The workflow section explicitly details the steps: Read the primary tool file ( ), follow the Cookbook to determine the tool, and execute the tool. The agent successfully executed an command in a forked terminal, demonstrating its ability to run raw CLI commands in a new context. When prompted to summarize the codebase, the agent correctly identified and utilized the and files, showing it understood the environment structure. The agent successfully handled context-aware summarization by using the file, which includes prior conversation history, proving its ability to maintain context across forks. The final structure of the skill is highly modular, separating concerns into directories like , , , , and .

Context: The video is a technical tutorial by IndyDevDan demonstrating the process of creating a custom, reusable agent skill within an agentic coding framework, likely related to the Cursor or similar AI-powered IDE environments. The specific skill being built is named "Fork Terminal Skill," designed to allow an agent to spawn new terminal windows to execute commands or access documentation relevant to the current task, which requires complex orchestration between different LLMs like Claude, Codex, and Gemini.