A new workflow guide from The Rundown AI outlines how creators can use Higgsfield with Claude Code or Codex to turn short-form video production into a repeatable terminal-based system. Rather than generating isolated clips, the setup is designed to help users organize campaigns, save outputs, track feedback, and refine prompts over time.
The guide is aimed at creators, marketers, agencies, and other operators who want a more structured way to test video ideas without rebuilding the same prompts for each new attempt. It also serves as an example of how CLI tools, reusable skills, and feedback loops can be combined inside a local project folder.
At the center of the system is a normal project directory that becomes a video workstation. Inside it, Claude Code or Codex can create campaign folders, generate videos with Higgsfield, store the results, and keep a record of what worked and what did not. The setup is meant to make production easier to repeat and improve, rather than forcing users to start from scratch every time.
The guide says the workstation can be used to create a campaign, generate two videos through the terminal, save the outputs, and then improve the prompts based on feedback. Over time, those steps can be converted into reusable skills so the process becomes more predictable and efficient.
To get started, the guide says users need an agent that can operate inside a local project folder, a Higgsfield account, access to the Higgsfield CLI, and an idea for a brand, product, or campaign to test. It also recommends reviewing prompts before any generation so credits are not spent blindly.
The guide advises against immediate automation. Instead, it recommends using the system manually for about five days or five campaigns so the agent has real examples to learn from. That early stage is intended to help Claude identify patterns based on actual feedback, rather than making assumptions about creative preferences.
After that learning period, the guide suggests setting up a daily automation. In the sample workflow, the agent would review previous campaigns and feedback, spot patterns in what performed well or poorly, propose a new campaign idea, create the folder, generate a set of videos, save the outputs, update a tracking file, and recommend an improvement to the generation skill.
Even then, the guide says the system should ask for approval before using Higgsfield credits unless the user explicitly enables automatic generation. The recommendation is to automate the review, ideation, and improvement steps first, then decide later whether full daily generation should run without human signoff.
The broader idea behind the guide is to move short-form video work from ad hoc experiments to a repeatable process that can scale from the terminal. By pairing Higgsfield with Claude Code, the workflow tries to give creators a structured way to test ideas, capture lessons, and continuously refine how prompts and campaigns are produced.
The guide’s takeaway is straightforward. Build the workstation, generate a small number of videos, save the outputs, convert the workflow into a skill, use feedback to improve it, and only then automate. In The Rundown AI’s framing, that sequence is what turns one-off video generation into an ongoing production system.