Substage

Want more ideas for what you can do?

For most commands, we highly recommand small, fast models such as GPT 4.1 mini. Substage is intended to be used for quick individual operations, such as file conversion, and with small AI models, this can be a quick and snappy experience.

As soon as you increase the complexity of your request, things can get more unreliable. We recommend that more complex requests are only done by developers and tech-savvy users who understand the Terminal commands that are generated.

In order to understand Substage's further limitations, it's worth reviewing how it works under the hood:

• Your prompt is converted to a Terminal command by an LLM.
• A small amount of context is sent to the LLM, including the names of the files you have selected, and most importantly, the file extensions.
• This will result in a command that will run locally on your Mac, at which point the LLM is no longer involved. Whether it's a video conversion, a zip operation or a Word doc to RTF conversion, the LLM doesn't seem the contents of these files.
• However, there is one exception to this: Substage also uses an LLM to summarise the output of the Terminal command. If the output of the command was the entire content of a text file for example, then the LLM can summarise this content.

Given that it's primarily a one step process, it's important to understand that the following will not work:

• "Organise this folder": Not possible, because the LLM isn't aware of the entire content on your computer, so wouldn't know where to start - this requires a lot more than a single Terminal command. However, if you have a specific idea of what you need, such as "put all mp4 files in a videos folder", then it works great.
• "Put this in that folder I was working in yesterday": the LLM doesn't have a history of what you've been working on in the past, and only knows the names of the files you currently have selected.
• Have a conversation with AI: Not possible, because each command is isolated. We have a feature request to allow some back and forth, but it's not yet implemented.

In addition, more complex operations aren't entirely recommended, for example if you describe multiple steps in a single prompt. It's best to describe each step separately, and run them one at a time. For example:

• "Convert all the mp3 files into aac, put them in a folder called 'audio' along with the originals, in separate sub-folders, and zip them all up": While this may work, it can be risky, since it's easy for the AI to get a tiny detail wrong. For most commands, we highly recommand small models such as GPT 4.1 mini since it's nice and snappy, but in this instance (if you insist on trying it!), we'd recommend using a larger model such as GPT 4.1 or Claude 3.7 Sonnet.

We don't currently have integration with AI providers beyond the described Terminal command conversion process. So, for example, you can't ask Substage to:

• "Transcribe this podcast audio file to text using Whisper" - not currently possible, but we're exploring ways to add this.
• "Rename these images based on their contents" - Substage currently has no mechanism for image content analysis, and would be a multi-step process that isn't currently supported.
• "What's this a picture of?" - As above, Substage currently has no mechanism for image content analysis.
• "Make a jpg of a pony" - Substage doesn't integrate with any image generation service.

And finally, a quick fire round of things that won't work in Substage:

• "Download all images from <some url>" - This would likely require a multi-step process. While it's conveivably possible to write this in a single line Terminal command, LLM aren't likely to get it right.
• "Install <some package> using homebrew" - We disallow installing new software for safety reasons.
• "Solve world hunger" - That would be nice.