Agentic coding workflow with local OpenSuperWhisper input

Kun walks through his terminal workflow for a coding assistant, using Claude Code for most demonstrations while comparing Codex CLI, Pi and OpenCode. He prefers WezTerm for its Lua configuration and Windows, Mac and Linux support. tmux provides persistent sessions and parallel panes; Neovim handles keyboard navigation and editing.

The tutorial separates global preferences from project memory. Kun keeps his global file short because it enters every session, then moves conditional instructions into skills that agents load when needed. He warns that downloaded skills can introduce security risks or worsen results, citing his own evaluation rather than treating popularity as evidence of quality.

OpenSuperWhisper supplies local AI voice input by running Whisper on the user's machine. Kun also demonstrates an initial prompt for project vocabulary. This local transcription step does not establish that the coding models run locally.

For planning, lavish creates HTML artifacts with annotations and selectable decisions. Kun's No Mistakes pipeline reviews and tests changes in an isolated worktree, records evidence and prepares a pull request. He still assesses the results before merging. gnhf loops over longer tasks with token or iteration caps and stop conditions. Treehouse manages reusable worktrees, while Firstmate coordinates parallel agents across repositories. The workflow depends on tools such as GitHub, and Kun's efficiency comparisons are reported benchmarks.