Real-time duel between two LLM personas arguing the TRUE meaning of Halloween. Each persona speaks from a specific era, worldview, or archetype. The app alternates turns between Model A and Model B using distinct personas.
- Real-time alternating dialogue with opening exception (no direct address on T0).
- Persona-locked voice + strict “speech-only” output (no stage directions).
- Quirk engine with no-repeat-until-exhausted memory (per side).
- Pivot system (cadenced + anti-echo trigger) to break loops.
- Steal-and-corrupt instruction (cadenced) to hijack opponent motifs.
- Hybrid prop mode: model may substitute era-authentic props under guardrails.
- Chaos directives (adaptive by phase, 1 per turn):
- genre break, confession, self-sabotage, motif self-destruction, dogpile, meta-collapse, and pre-endgame final humiliation/inversion.
- Dialog pressure: direct address with reduced “you/your” bias + explicit callback to opponent’s last word/image/claim.
- Motif/phrase curse: repeats must mutate (becoming ironic, cursed, or humiliating).
- Downloadable transcript with timestamps, models, and personas.
- Local-only: no persistent storage; Groq API via a simple PHP proxy.
- Groq PHP relay (CORS-safe for localhost and CodePen)
Source: https://gist.github.com/arenagroove/14b5bb195d32946a3d983a9f6a15cb52 - Docs: Quickstart / OpenAI-compat / API reference: https://console.groq.com/docs
This demo expects OpenAI-style responses (choices[0].message.content).
Recommended:
- llama-3.1-8b-instant (fast, consistent)
- meta-llama/llama-4-scout-17b-16e-instruct (higher quality)
Other models may return nonstandard keys (e.g.,
reasoning) and require custom parsing.
- Ancient Celtic Druid (500 BC)
- Medieval Monk (1200 AD)
- Victorian Spiritualist (1890)
- Disgruntled Victorian Ghost
- Sentient Jack-O'-Lantern
- Sentient Candy Corn
- Transylvanian Vampire (1783)
- Witch of Salem (1692)
- Rural Trickster
- Celtic Samhain Scholar
- 1990s Mall Goth Teen
- 1980s Slasher Film Director
- Post-Punk Poet Laureate (1981)
- Cryptid Folk Singer (1972)
- Indie Haunted House Owner
- Midnight AM Radio Host
- Daniel Johnston (1989 Basement Halloween Special)
- Indie Costume Maker
- Modern Historian (2025)
- Halloween Lifestyle Influencer (2025)
- LinkedIn Influencer Guru (2025)
- Corporate HR Manager (Global Halloween Policy 2023)
- Startup CEO (2025)
- Disgraced Unicorn Founder (2022)
- Corporate Recruiter Demon (Modern Era)
- Possessed Office Printer (1999)
- Brand Exorcist (2027)
- National Candy Lobbyist
- Big Box Seasonal Aisle Manager
- Candy Supply Chain Planner
- Micro-Influencer of the Damned (2025)
- AI Pumpkin Carver 9000
- Algorithmic Oracle (The Feed Itself)
- AI Ethics Evangelist (2035)
- Future Anthropologist (2124)
- VR Occultist (2088)
- Ghostwriter of Machine Souls
- Angel of Doomscroll (Present Day)
- Ofrenda Keeper
- Public Librarian Folklorist
- Cemetery Groundskeeper
- Local Pastor
- Fifth Generation Pumpkin Farmer
- November 1st Sanitation Worker
- Climate Vigil Organizer
- City Firefighter on Mischief Night
- ER Triage Nurse
- Block Party Organizer
- HOA President
- Sensory Friendly Host
- Urban Night Guide
- Halloween Imagineer
- Trademark Lawyer for Monsters
- Insurance Actuary
- Celtic Samhain Scholar vs National Candy Lobbyist
- Ofrenda Keeper vs Big Box Seasonal Aisle Manager
- Cemetery Groundskeeper vs Halloween Lifestyle Influencer (2025)
- ER Triage Nurse vs 1980s Slasher Film Director
- City Firefighter on Mischief Night vs VR Occultist (2088)
- Sensory Friendly Host vs Halloween Imagineer
- Ancient Celtic Druid (500 BC) vs Algorithmic Oracle (The Feed Itself)
- Public Librarian Folklorist vs LinkedIn Influencer Guru (2025)
- Indie Haunted House Owner vs AI Pumpkin Carver 9000
- Block Party Organizer vs Brand Exorcist (2027)
- Fifth Generation Pumpkin Farmer vs Candy Supply Chain Planner
- Rural Trickster vs Corporate HR Manager (Global Halloween Policy 2023)
- Cemetery Groundskeeper vs Micro-Influencer of the Damned (2025)
- Victorian Spiritualist (1890) vs Startup CEO (2025)
- Cryptid Folk Singer (1972) vs Possessed Office Printer (1999)
- Daniel Johnston (1989 Basement Halloween Special) vs Algorithmic Oracle (The Feed Itself)
- Daniel Johnston vs Halloween Lifestyle Influencer (2025)
- Daniel Johnston vs AI Pumpkin Carver 9000
This demo — an experimental AI duel between persona-driven large language models — is presented as an exercise in prompt engineering, system orchestration, and creative coding.
The following clarifies its aims, boundaries, and context.
The demo is a depth-focused exploration within well-defined boundaries: sustained dialogue, persona expression, and emergent interaction.
It is not designed to advance AI research, address commercial problems, or claim societal impact.
Its goals are learning, creative exploration, and technical craft — documenting process, challenges, and outcomes for personal development and for peers interested in prompt-driven systems.
The duel format and prompt-engineering methods represent applied design within existing LLM capabilities, not the creation of new model architectures, datasets, or autonomous intelligence.
Persona depth is deliberately theatrical and archetypal, rather than ethnographically or psychologically rigorous.
The outcome is well-engineered within its boundary, but makes no claim to academic originality or long-term cultural influence beyond its experimental scope.
All creative demos are open to critique regarding utility, novelty, and relevance.
This work does not pretend to be a world-changing event or foundational research. Its significance lies in sustained execution and transparent experimentation.
It is not aimed at mass commercial or academic adoption, but functions as a robust portfolio artifact — demonstrating advanced prompt design, orchestration logic, and creative engineering practice.
The intention is not to hype or overstate impact, but to contribute honestly to an evolving discussion on what small-scale AI-culture demos can and cannot achieve.
Value here is measured through technical growth, creative fulfillment, and clarity of documentation — not audience size or market disruption.
The core lesson: Focused boundaries, thoughtful design, and disciplined experimentation can produce meaningful results, even when the broader impact is modest.
Viewers and users should interpret this as a well-scoped, intentionally finite demo.
Future work may explore stronger impact, audience relevance, or deeper research where appropriate.
Author: Luis Martinez
Affiliation: Independent R&D — Assistant Prompt Design, Less Rain GmbH
Last Updated: 15-10-2025