Strategy AI Creative Studio
An internal AI creative studio built inside Discord, exposing image, video, audio, 3D, and brand-locked generation to 70+ users at Strategy through a single bot and a tier-based channel architecture.
Right-click any image to open it full size in a new tab.
Strategy’s creative team was using generative tools constantly, but the experience was fragmented: different platforms, different logins, no shared history. Every session started from scratch.
Strategy AI Creative Studio fixes that. One Discord-native interface to image, video, audio, 3D, and brand-locked generation, used by 70+ people across the company.
Why Discord
The team already lived there. Outputs land in shared channels, so prompts and references compound into a team library. Anyone can browse what others have made, fork a prompt, or pull a reference URL into their next command. Discord’s slash commands, modals, and threads also meant the right surface was already in place.
Channels as models
The core design move: each channel maps to a specific model and use case. Users don’t pick a model. They go to the channel that produces what they want, type a slash command, and describe the output. The bot handles model routing, parameter setup, queue management, and delivery.
Channels are organised in tiers. Tier 1 is brand-locked output (#strategy-neon, #strategy-premium, #maxi-generator) using hidden reference images and locked prompt prefixes server-side. You cannot accidentally go off-brand. Tier 2 is general image generation through FLUX 2, FLUX 2 Pro, Nano Banana 2. Subsequent tiers cover image editing, video, audio, 3D, and experimental apps.
The architecture is legible at a glance. Channel name tells you what it produces. No settings panel, no model picker. The complexity is hidden behind a discipline of channel design.
Apps that compose multiple stages
A few channels run multi-step pipelines that look like single commands. The Maxi Poster app runs scene generation, headline overlay, automatic logo placement, and Topaz upscale as one workflow. The user types one command with a scene description and a headline. About 60 seconds later, a finished 9:16 portrait poster lands in the channel.
Each stage has a single job, fails predictably, and can be swapped out independently when a better model becomes available.
Operational design
The interesting engineering problems weren’t model integrations, they were operational. A queue system that lets six people hit the bot simultaneously without job collisions. Async polling against the FAL API so users get progress updates rather than silent timeouts. Auto-deploys from GitHub on every push. A /ask command that routes anyone to the right channel with a suggested command.
Built in Python on discord.py and aiohttp, hosted on Render, with FAL as the primary model provider. Models running through it: FLUX 2 and FLUX 2 Pro, Nano Banana 2, Kling V3 Pro and O3, Seedance, Veo, Grok Imagine, MiniMax for speech and music, ElevenLabs, Hunyuan 3D, Bria, Topaz. Eleven services behind one consistent grammar.
Outcomes
70+ active internal users at peak. Tens of thousands of assets generated. Campaign production cycles that depended on stock licensing and external shoots now happen inside the studio in hours. The patterns I proved here (channel-as-model, locked brand prefixes, multi-step apps as single commands) became the operating model for how Strategy thinks about internal AI tooling.