Demo: Songbird GPT (boogie/song/colorful engines) in Action

Building a custom GPT is one thing. Watching it actually work, unedited, on real input is another. This demo does the second thing: a first, unpolished run of Songbird GPT, a small custom GPT built to generate music-video ideas, tested live on a Sanskrit devotional chant with no cleanup of what comes back.

What Songbird GPT does

Songbird GPT is a simple tool with a handful of modes rather than a sprawling feature set. Boogie handles dance videos, song turns lyrics into music-video sequences, and colorful applies an iPhone-style hyper-color look. There is also an unreal mode, and a small holiday switch where typing "mantra" swaps the theme to Krampus or Santa Claus depending on the season. The tool is still very much a work in progress, and the modes themselves are deliberately narrow rather than trying to do everything at once.

Why the sequencing numbers matter

One detail gets specific attention in the demo: the numbers the tool adds to each generated prompt. Most text-to-image tools ignore those numbers outright, but they serve a practical purpose anyway, since image generators like Midjourney typically name output files using the first few words of the prompt. Without a number at the front, there is no reliable way to know which image comes first, second, or third once a batch of stills comes back. Adding a sequence number keeps the clips in order by controlling how the filenames sort, which matters when the whole point is building a video from ordered B-roll.

Testing it live: song plus colorful

For this demo, the input was Gaia's Mantra, a Sanskrit devotional chant repeated as a mantra. Running it through the song and colorful settings together, with no heavy editing of the returned prompts, is meant to show exactly what the tool produces right now, not a polished best-case result. That is a deliberate choice: rather than refining prompts and layering in custom footage, as would normally happen in a finished video, this demo is presented as close to 100% raw B-roll straight from the tool.

How this fits the larger workflow

The plan going forward is to use Songbird GPT to generate B-roll, then refine the prompts and add custom footage on top of what the tool returns, rather than relying on the raw output as a finished product. Like most AI tools, the expectation is that it will need repeated refinement over time rather than working perfectly on the first pass. Showing the unrefined version first sets an honest baseline for what improves as the tool and the workflow around it develop.

Key takeaways

• Songbird GPT is a custom GPT with narrow, mode-based functions (boogie, song, colorful, unreal, and a holiday-themed switch) rather than one general-purpose tool.
• Sequence numbers added to each prompt solve a real problem: most text-to-image tools sort output by filename, and numbering keeps generated clips in the intended order.
• This demo intentionally shows raw, unedited tool output rather than a refined final product, to give an honest look at where the tool currently stands.
• The typical workflow is to use the tool's output as B-roll, then add custom footage and refine prompts further rather than using the raw results as-is.

Try it yourself

If you build video content and want to see how Songbird GPT works, a link to the page explaining how to use it is available through the Humanitarians AI Songbird page. This is a useful watch for creators curious about building their own lightweight custom GPT tools for generating B-roll and music-video ideas.