AI Adoption · 2023–2026
Building an AI-Native Creative Function
- AI
- Branding
- Content
Overview
Over the last few years I have been building AI into the creative function at Filecoin Foundation as working infrastructure. The push came from a practical place - the resources got smaller and the work didn’t - and the answer was an adoption programme: training, building and integrating until the brand’s own knowledge lives in every system we use, from strategy research through to the video edit. This page explains what that looks like in practice.
Approach
Learning by doing. Every tool in the stack started as an experiment: run a real task through it, look hard at what the first pass gets right and wrong, then refine the tool itself - the material it’s trained on, the instructions it works from, the examples it’s shown - until the output is reliable. The judgment stays human - each tool earned its place in the workflow by the eventual quality of its output.
It’s a lot like working with an enthusiastic intern: the raw ability is there from day one, and the quality of the results comes down to the conditions you create.
Agentic brand strategy
Brandy is a brand strategist agent I built on Perplexity and Claude. Training it meant grounding it in established brand-strategy literature - Aaker, Ries and Trout, Kapferer, Neumeier - so when it analyses a brand it applies the same human frameworks a strategy consultancy would, with sources you can check. In the Filecoin repositioning that worked like this: Brandy assessed the brand against those frameworks, then mapped Filecoin, its storage competitors and the major chains onto a positioning chart - how optimistic each brand’s story is on one axis, how human- or technology-focused on the other. The map showed an open space: almost nobody in web3 was telling an optimistic, human-focused story. I pushed back on placements where I disagreed, it corrected or made its case, and the version we settled on became the creative brief. A second agent - a copywriter trained on the brand’s voice - drafted the expression for each candidate positioning, and Brandy reviewed the drafts against the strategy. Every agent in the stack follows that pattern: write the thinking down, train it in, and use the agent as a sparring partner rather than a vending machine.
The event archive, made useful
Every event produces hours of recorded talks that are usually watched once and shelved. Maple, Donut and Dulce - knowledge bases we build after each event - turn them into something people can keep using. The process is simple enough to copy: record every session, transcribe the recordings with Whisper, clean up the transcripts, then load them into a Perplexity Space or NotebookLM notebook along with the speaker decks and supporting documents. For FIL Dev Summit 8 in New York, the finished notebook went to every attendee as a link. Anyone with it can ask the summit questions and get answers grounded in what was actually said on stage, read the generated briefing docs and FAQs, or listen to a podcast-style audio overview of the week. Aside from significantly reducing production overhead and costs - allowing one person to capture content in multiple rooms with cheap audio recorders - the event documentation becomes a reference that keeps answering questions after everyone has flown home.
Teaching the tools the visual language
Space punk - the generative visual language from the Filecoin rebrand - stays consistent because its prompts are governed like any other brand rule. The exact prompt language that produces the look is written into the brand guidelines, sitting alongside the colour palette and the type system, and it runs on Midjourney, so anyone on the team can generate new imagery that matches the hundreds of pieces before it. The same thinking applies to the guidelines themselves: the brand book and visual identity are loaded into Claude design, allowing ecosystem participants to self-serve in the tools they want to use.
Using AI to ship code and video
The stack reaches past design into production. Website updates and fixes for filecoin.io go through Claude Code: I describe the change, review what it builds, and merge it - the same flow as working with a developer, without waiting on engineering availability. This portfolio was built the same way. On the video side, the pipeline runs on Descript, where Underlord handles the rough-cut edit passes - the assembly work that used to consume production days - leaving the finishing decisions to a human.
Outcomes
The function does more per person than it ever has. The brand holds its consistency, events are documented and stay useful long after they end, the sites keep shipping and the publishing cadence never broke - all through a period when the workload was built for a bigger team. Because each system is documented and trained rather than living in someone’s head, the systems transfer: a new team member inherits working tools, and the whole approach could move to another brand tomorrow. This is how the function runs now, and how I would build one from scratch.