Contributing to Ghost Pitch Engine

3 roles, 4 files, no build tools.

File Structure

index.html          ← Config, state, DOM, TTS, crowd detection, event wiring
js/llm.js           ← LLM framework (shared — don't break it)
js/ghosts.js        ← Ghost generation (Dev 2's zone)
js/show.js          ← Show runner, judging, Boo MC, OVERDRIVE (James's zone)
CLAUDE.md           ← Agent instructions (read this if using Claude Code)
HANDOFF.md          ← Architecture + work split + secrets
CONTRIBUTING.md     ← This file

Roles

Dev 2: Ghost Generation & Pitch Presentation

Your file: js/ghosts.js

What to do:

1. Edit GHOST_PROMPT to make ghost generation prompts more creative
2. Add MORE models to CONFIG.featherless.models.corpus in index.html — llmFanOut hits all of them in parallel (12,000+ available at featherless.ai/models)
3. Make the pitch card visually pop when a ghost presents (edit CSS in index.html)
4. After generation, index pitches into Needle (see HANDOFF.md)
5. Keep FAILSAFE_GHOST ("Der Speisekarten-Geist") as the last ghost

Don't touch:

• js/show.js — judging, Boo, crowd, OVERDRIVE
• js/llm.js — LLM framework (use it, don't change it without asking)

Test your changes:

git push origin main
open https://hackathon-berlin-mar-2026.westover.lol

Business: Ghost Judge Backstories

You edit CONFIG.judges in index.html — the backstory strings.

Each judge has:

{
  key: 'carla',                    // don't change
  name: 'Karla von Strategos',    // ghost name (can edit)
  voiceIdx: 2,                    // don't change
  backstory: 'You are the ghost of Carla Schneider. In life...',  // EDIT THIS
}

Guidelines for backstories:

• Start with "You are the ghost of [Real Name]."
• Include how they died (funny/dramatic)
• Include what they judge (business viability, technical depth, product craft)
• Include a catchphrase
• Keep under 200 words — these go into LLM prompts

Also edit:

• BOO.backstory in js/show.js — the MC ghost's personality
• Ghost names in CONFIG.judges[].name — make them spooky/funny

James: Judging + Crowd + OVERDRIVE

My file: js/show.js

My zone:

• judgeGhost() — judge deliberation with Boo + core + chaos judges
• booJudge(), booSpeak(), summonChaosJudges()
• decideBuild() — BUILD/NO BUILD via Boo
• overdrive() — build + deploy pipeline
• runShow() — main flow orchestration
• Needle integration for judge RAG context

Also own in index.html:

• listenToCrowd() — Web Audio FFT crowd detection
• Crowd detection config (CONFIG.crowd)

Config Block (in index.html)

Key	What	Who edits
featherless.apiKey	API key (expires 2026-03-17)	nobody
featherless.models.corpus	Models for ghost generation	Dev 2
featherless.models.judges	Models for judge LLM calls	James
needle.apiKey	Needle RAG API key	nobody
crowd.*	FFT detection params	James
tts.*	Speech rate/pitch defaults	anyone
judges[]	Judge personas + backstories	Business

Rules

1. Don't break the flow — the show runs top to bottom in runShow()
2. Keep it fast — 2.5 minutes total for the presentation
3. TTS rate >= 1.1 — ghosts talk fast
4. Judge verdicts = 1 sentence — maxTokens: 60
5. Test at the live URL — local file:// won't work (needs HTTPS for mic)
6. Use the LLM framework — never call fetch() to Featherless directly, use llmCall/llmJSON/llmFanOut/llmStream