Building an AI-Native Creative Function

Overview

Over the last few years I’ve been building AI into the creative function at Filecoin Foundation as working infrastructure. The interest was in finding practical ways to apply it to real problems for immediate results, rather than expecting magic out of a black box. That meant training, building and integrating until the brand’s own knowledge lives in the systems we use, from strategy research through to the video edit. This page is what that looks like in practice.

Approach

Everything in the stack started as an experiment. Run a real test through it, look hard at what the first pass gets right and wrong, then go back and iterate on the tool. Give it new data, change its instructions, give it better examples, and above all give it direct feedback on its output. Then run it again. That loop is the whole method.

Nothing goes out unseen by human eyes and unvetted by human thought. The machine isn’t there to replace the thinking, it’s there to accelerate it, and the key to using these tools well is creating the conditions for that to happen.

The event archive, made useful

Open-source ecosystem events are chaotic — people flying in from everywhere, a few days to get through their objectives, and always a session on the other side of the building they can’t get to. We put a cheap audio recorder in every room, ran the audio through Whisper, and had AI clean the transcripts afterwards, which it handles better than the transcription does, because it knows from context that “Falcon” means “Filecoin.” The cleaned transcripts went into a tool people could question — what did I miss yesterday, what’s worth following up in my area, what did she say about that on Tuesday. It replaced the video documentation and the note-taking, and it meant people could stay in the conversation they were already in.

Brand strategy and copy

The goal was to reposition the Filecoin brand to resonate better in a changing world — and doing that meant crafting two tools by hand before bringing them together.

One was a brand strategist trained to apply real frameworks rather than guess words — Neumeier’s five disciplines, Ries and Trout on positioning, Aaker, Kapferer’s prism, Byron Sharp. Human frameworks applied to the problem, not vibes.

The other worked backwards. Starting from 80,000 Hours and its account of large-scale problems facing humanity, I established what the relevant problems were first, then reasoned forward to where Filecoin could genuinely be applied to each.

Refinement was a loop in both cases. Take the output, edit it, hand the edit back and ask it to retain what we learned. Other rounds went through the team — suggest twenty concepts most important to the Filecoin brand, give that to people with institutional knowledge, have them rank it, annotate it, flag what was missing, and feed that back.

What came out was a brand strategy that could carry the scale of what Filecoin was actually trying to do, and then website copy, blogs and social content built off it. Not finished on the first pass, but eighty percent there — which meant the team pulled up usable drafts in seconds and spent their time on the detailed work humans are better at.

These aren’t a finished project. They’re still in use — when a new product comes up I can put it through positioning analysis, generate options and prune them down to something useful, and keep the copy current as more of it gets written. They’re working tools the team’s day-to-day output depends on.

Teaching the tools the visual language

The biggest danger to brand coherence now is probably AI tools. People who aren’t qualified to make aesthetic decisions can generate something they think looks good — and it may well look good — but it can also break the guidelines. Rather than try to stop people making slop, we thought it better to put something in their hands that makes good work.

So we built an illustration style in Midjourney, through the same loop as everything else: generate, pick the best, feed it back, generate again. What came out of that was prompt language, and prompt language went into the brand book as an element of the visual identity, alongside everything else. Then rather than gatekeep it, we shared it with the whole ecosystem, so anyone building on Filecoin could generate work that held together with ours.

Human creativity, AI accelerated

Design reviews happen as a conversation. Two of us on a call, screen shared, working through a page from top to bottom: why that sits there, do you prefer this to that, what was the reasoning. We make the list of changes verbally as we go.

Then the transcript goes into Claude Code, which executes what we decided. Two designers having a real discussion, learning from each other, rather than each talking to a robot separately.

Using AI to ship code and video

Websites used to pass through design between content and development. Now a Figma prototype, or even just a wireframe or sketch, can go straight into Claude Code. We’ve trained projects on our existing sites, so changes can be made on the fly and the code stays consistent across the whole thing.

The first instinct now is to see what it comes up with and build from there. That’s a starting point, not a finished one, and the code still gets checked. It’s enabled designers to work directly on the finished product and freed up developers for the hard work that requires human attention.

Video works the same way. Descript goes through long recordings to find the specific moments and put together a rough assembly, and the detailed work stays with a person where it needs to.