MAX CAPACITY

I built a fully autonomous D&D campaign engine that plays itself. Three AI players. One AI dungeon master. CrewAI orchestrating. DeepSeek running the brains. Python handling all the bookkeeping. Cost: about $0.20 per session. It runs daily, updates a blog with session recaps, and maintains a wiki of every NPC, location, and faction the party encounters. The hard part wasn't getting AI to play D&D. The hard part was getting it to play a *campaign*. Multiple sessions. Multiple adventures. Characters that level up. A world that remembers. Some of what I had to solve: When an adventure completes (3 sessions each), Python detects it, loads the next adventure file, resets the state, and keeps going. No human needed. 20 adventures across 4 acts, all mapped out. Milestone leveling fires automatically between adventures. HP by class hit die, new class features, spell slots, cantrips. A level 3 fighter knows he has Action Surge. A level 1 fighter doesn't. The agents' prompts update to match. The wiki was a disaster at first. 155 entries in 10 sessions, mostly junk like "the door" and "cast fireball." Rewrote the Wiki Keeper with strict filtering: named NPCs, named locations, named factions, unique items only. Session 1 now produces ~15 clean entries. Story summaries were ballooning. 1,500+ words by session 10, eating context window and burning tokens. Now they get condensed to key plot points when an adventure wraps. The DM kept skipping rest periods. Six sessions, no long rest, party running on fumes. So when a new adventure starts, Python injects a mandatory rest and resupply block into the DM's opening prompt. Biggest lesson: you have to be very clear about what Python owns vs. what the AI owns. Session tracking, adventure progression, level data? Python. The AI touches none of it. The moment you let the DM agent "help" manage state, it will hallucinate session counts and break everything. Still early. Need to test multi-session persistence over weeks, not days. But the foundation is solid and the sessions are genuinely fun to read. Fully autonomous AI D&D. $0.20 a session. The future is weird.

I built a fully autonomous D&D campaign engine that plays itself. Three AI players. One AI dungeon master. CrewAI orchestrating. DeepSeek running the brains. Python handling all the bookkeeping. Cost: about $0.20 per session. It runs daily, updates a blog with session recaps, and maintains a wiki of every NPC, location, and faction the party encounters. The hard part wasn't getting AI to play D&D. The hard part was getting it to play a *campaign*. Multiple sessions. Multiple adventures. Characters that level up. A world that remembers. Some of what I had to solve: When an adventure completes (3 sessions each), Python detects it, loads the next adventure file, resets the state, and keeps going. No human needed. 20 adventures across 4 acts, all mapped out. Milestone leveling fires automatically between adventures. HP by class hit die, new class features, spell slots, cantrips. A level 3 fighter knows he has Action Surge. A level 1 fighter doesn't. The agents' prompts update to match. The wiki was a disaster at first. 155 entries in 10 sessions, mostly junk like "the door" and "cast fireball." Rewrote the Wiki Keeper with strict filtering: named NPCs, named locations, named factions, unique items only. Session 1 now produces ~15 clean entries. Story summaries were ballooning. 1,500+ words by session 10, eating context window and burning tokens. Now they get condensed to key plot points when an adventure wraps. The DM kept skipping rest periods. Six sessions, no long rest, party running on fumes. So when a new adventure starts, Python injects a mandatory rest and resupply block into the DM's opening prompt. Biggest lesson: you have to be very clear about what Python owns vs. what the AI owns. Session tracking, adventure progression, level data? Python. The AI touches none of it. The moment you let the DM agent "help" manage state, it will hallucinate session counts and break everything. Still early. Need to test multi-session persistence over weeks, not days. But the foundation is solid and the sessions are genuinely fun to read. Fully autonomous AI D&D. $0.20 a session. The future is weird.

I'm teaching AI to curate my art, not create it. I have 20 years of raw VHS glitch footage sitting on hard drives. Hundreds of thousands of frames. Somewhere in there are incredible stills — moments where a recognizable subject is visible through the glitch artifacts. Finding them manually would take months. So I'm building a two-phase curation pipeline with Claude Code. Phase 1 is a local Python script that scores every frame on structure and glitch intensity — Claude wrote it, I tune it by reviewing the results and giving feedback. It processes 30 images per second at zero cost and eliminates 90-97% of frames as uninteresting. Phase 2 sends only the survivors to Claude's vision model for final scoring. The key breakthrough was multiplicative scoring: a frame needs BOTH a recognizable subject AND strong glitch artifacts to pass. Took 11 iterations to get here. Each round I'd find a new failure mode — scanlines fooling the filter, dark atmospheric frames getting rejected, abstract color blobs sneaking through — and we'd add a new metric to catch it. Currently processing 300k frames across multiple sources. Targeting over a million for calibration. The frames that survive are going to become a new collection.

I'm teaching AI to curate my art, not create it. I have 20 years of raw VHS glitch footage sitting on hard drives. Hundreds of thousands of frames. Somewhere in there are incredible stills — moments where a recognizable subject is visible through the glitch artifacts. Finding them manually would take months. So I'm building a two-phase curation pipeline with Claude Code. Phase 1 is a local Python script that scores every frame on structure and glitch intensity — Claude wrote it, I tune it by reviewing the results and giving feedback. It processes 30 images per second at zero cost and eliminates 90-97% of frames as uninteresting. Phase 2 sends only the survivors to Claude's vision model for final scoring. The key breakthrough was multiplicative scoring: a frame needs BOTH a recognizable subject AND strong glitch artifacts to pass. Took 11 iterations to get here. Each round I'd find a new failure mode — scanlines fooling the filter, dark atmospheric frames getting rejected, abstract color blobs sneaking through — and we'd add a new metric to catch it. Currently processing 300k frames across multiple sources. Targeting over a million for calibration. The frames that survive are going to become a new collection.

I guess many artists that were always in Eth that had total disdain, made fun and looked over the shoulder to other artists that trusted on Tezos ecosystem, are and will arrive to this blockchain, because they are finally realizing most of their collectors were not collectors, just cryptobros/speculators that used them as 'tokens', and now that they left, after making profits at their expense using their art as mere 'product' (and artists did MILLIONS), they are realizing that the CULTURE is still alive in the 'minor blockchain'... Some pave the way, others enjoy it (as usual). You are welcome.🫴 Let me repeat what I said in 2021: Eth: People involved in Crypto stumbling upon Art. Tz: People involved in Art stumbling upon Crypto. Money destroys culture. Art creates it. Let's work.✊