Guides The Rundown AI published jun 1, 2026
In this guide, you will learn how to set up a Codex project that lets you test, generate, and improve short-form video with Higgsfield from your terminal. Instead of making one-off clips, you will create a video workstation with campaign folders, saved outputs, reusable skills, and a feedback loop.
You will build a Higgsfield video generation workstation inside a normal project folder. The folder gives Claude Code or Codex a place to create campaigns, save generated videos, track feedback, and turn the process into reusable skills.
By the end, you will have a working setup that can create a campaign, generate two Higgsfield videos from the terminal, save the outputs, improve the prompts from your feedback, and eventually run the same process on a schedule.
Use the system manually for about five days or five campaigns before you automate it. That gives Claude enough examples to learn from your real feedback instead of guessing your taste.
After that, ask Claude to create a daily automation:
Create a daily automation for this Higgsfield video workstation.
Every day, it should:
1. Review the previous campaigns and my feedback.
2. Identify patterns in what worked and what failed.
3. Come up with one new campaign idea.
4. Create the campaign folder.
5. Generate one set of videos.
6. Save the outputs and update tracking.md.
7. Suggest one improvement to the generation or iteration skill.
Before spending Higgsfield credits, show the planned campaign and commands for approval unless I explicitly turn on automatic generation.Do not automate blind generation on day one. Start by automating the review, idea generation, and skill-improvement loop. Once the campaigns are consistently good, you can decide whether daily generation should run automatically or wait for approval.
The reusable pattern is simple:
1. Create the workstation.
2. Generate two videos.
3. Save the outputs.
4. Turn the workflow into a skill.
5. Give feedback.
6. Turn the improvement loop into a skill.
7. Automate after the system has examples to learn from.That is how you move from one-off Higgsfield experiments to a short-form video system you can keep scaling from the terminal.
Start with a new folder for the workflow. Do not run this from a random downloads folder. The point is to give the agent one stable workspace where it can create campaign files, skill files, and tracking docs.
Use something simple:
mkdir higgsfield-video-workstation
cd higgsfield-video-workstationThen open Claude Code in that folder:
claudeYou can use Codex instead if that is your preferred agent. The important part is that the agent is operating inside the same folder where the Higgsfield workflow will live.
**Pro tip:** Use a dedicated folder for this. Skills and campaign tracking only become useful when the agent can find the same files again next time.
Open the Higgsfield CLI setup page and follow the three-part flow: install the CLI, sign in, and add the Higgsfield skills package to your agent.
Install the CLI:
npm install -g @higgsfield/cliThen authenticate:
higgsfield auth loginThe login opens a browser window and asks you to connect Higgsfield. Review the permissions before you approve access. The CLI needs to access generated content, uploaded media, workspaces, subscription plans, and credits so it can run and retrieve generations.
After that, add the Higgsfield skills package:
npx skills add higgsfield-ai/skillsWhen prompted, choose the Higgsfield generation skill and install it at the project level if this workstation is meant to stay self-contained.
Before you generate anything, ask Claude to check the setup:
Check whether the Higgsfield CLI is installed and authenticated in this project.
Do not generate anything yet. First tell me:
- which Higgsfield commands are available
- whether the Higgsfield generation skill is installed
- what command you would use for a low-risk test generation
- anything that might spend credits or require approval**Pro tip:** Make the agent show the planned command before it runs a generation. That one approval step saves you from wasting credits on the wrong duration, model, or prompt.
Now have Claude interview you about the brand and workflow. This is the setup step that turns a blank folder into a video generation workstation.
Use this prompt:
Interview me about my brand, audience, video goals, formats, and campaign structure.
Then set this folder up as a Higgsfield short-form video workstation.
Create:
- campaigns/
- skills/
- a brand guidelines file
- a campaign tracking template
- a README that explains how to run the workflow
Inside each campaign, use:
- input/
- output/
- drafts/
- finals/
- prompt-generation notes
- tracking.md
Do not generate videos yet.Answer the interview quickly. Claude should ask about the brand, audience, platform, video format, visual style, and what counts as a good clip. You are not writing a full creative strategy. You are giving the system enough context to make the first campaign coherent.
Once Claude creates the files, inspect the folder. If it creates too many layers, ask it to simplify. If it misses tracking, ask it to add a campaign tracking file before moving on.
**Pro tip:** Keep the first folder boring. A clean `campaigns/`, `skills/`, brand file, and tracking file is better than a giant production system you stop using after one day.
Now create the first campaign. This is the manual run that teaches the system what the reusable skill needs to do later.
Use a prompt like this:
Create a new campaign for [BRAND_OR_PRODUCT].
Use the brand guidelines and campaign template in this folder.
For this campaign:
- write a short campaign brief
- create two short-form video concepts
- turn each concept into a Higgsfield-ready prompt
- show me the commands before running them
- generate two videos with Higgsfield from the terminal
- save the outputs in the campaign output folder
- update tracking.md with the prompt, command, output file, status, and notesClaude should prepare the prompts and commands first. Review them before it runs anything. In the demo, Claude caught a practical constraint: Higgsfield's video model had a three-second minimum, so a two-second test needed to become a three-second generation.
When the generation finishes, check that the outputs are saved in the campaign folder, not just linked in the chat. The folder should include the campaign brief, prompt notes, generated drafts, and tracking file.
**Pro tip:** Do not turn the workflow into a skill until one full campaign has run end to end. The first run shows you what the skill needs to remember and what approval steps should stay manual.
After the first two videos are saved, ask Claude to package the working process as a skill.
Use this prompt:
Turn the campaign generation workflow we just completed into a reusable skill.
The skill should:
1. Read the brand guidelines.
2. Create a new campaign folder.
3. Ask for missing campaign details.
4. Write two Higgsfield-ready prompts.
5. Show planned commands before generation.
6. Generate two videos.
7. Save outputs in the campaign folder.
8. Update tracking.md.
9. Summarize what worked and what needs another pass.
Do not include one-time setup steps like installing the Higgsfield CLI. Assume the project is already configured.That last instruction matters. A skill should capture the repeatable work, not the setup you only do once. If the skill tries to reinstall tools every time, have Claude strip that out.
Test the skill on a second campaign or a different product angle. If it asks too many setup questions, point it back to the brand file. If it skips approval before generation, add that as a required step.
**Pro tip:** Separate setup from execution. The project README can explain installation. The skill should focus on creating campaigns, generating videos, saving outputs, and tracking results.
The first two videos are not the finish line. They are the first data point. The next move is to give Claude feedback and make the prompt-improvement loop reusable too.
Use this prompt after reviewing the videos:
Review these two generated videos against the campaign brief.
Here is my feedback:
- Video 1: [WHAT WORKED / WHAT DID NOT]
- Video 2: [WHAT WORKED / WHAT DID NOT]
Improve the prompts for both videos.
Generate the next versions.
Save the new outputs next to the originals.
Update tracking.md with what changed and why.Once that works, turn the iteration process into its own skill:
Turn this prompt-improvement process into a reusable skill.
The skill should:
1. Read a campaign folder.
2. Compare the campaign brief, original prompts, outputs, and feedback.
3. Improve the prompts.
4. Generate the next versions.
5. Save the revised outputs.
6. Update tracking.md with the prompt changes and feedback summary.This gives you two clean workflows: one skill creates new campaigns, and another improves existing campaigns. Keeping those separate makes the workstation easier to use.
**Pro tip:** Feedback should be specific. "Make it better" will not teach the system much. "The first frame needs a clearer product shot, the motion feels too slow, and the hook should happen in the first second" is useful.
Keep going
Members get all guides, courses, live sessions, and $1,000+ in partner perks.