How CultCodex Works

CultCodex is an independent fan archive — not an official Cult of Psyche publication. It indexes episodes, guests, lore, and topics from the @CultofPsyche and @PsychesNightmares YouTube channels to make the archive searchable and explorable.

Episode metadata (titles, dates, thumbnails) is pulled directly from the YouTube API. Transcripts are fetched from YouTube’s auto-generated captions where available.

1. Transcript acquisition

Transcripts come from YouTube captions where available. When captions are disabled, the audio is downloaded and transcribed locally with OpenAI Whisper (medium model). Either way, transcripts are stored as timestamped segments, which is what lets a search result or an Oracle citation jump to the exact moment in the video.

2. AI enrichment

Each transcript is processed by Anthropic Claude to extract structured data: episode summaries, guest identifications, topic tags, notable quotes, lore references, and content-type classification. The model is prompted with explicit editorial guidelines that require neutral, factual language.

3. Database import

Enriched data is imported into a PostgreSQL database (Xata) via Prisma. The import deduplicates people, topics, and lore entries by slug, so one guest with three display names still resolves to one profile.

4. Provenance badges

Every episode shows whether its data is “transcript-backed” (derived from a full transcript) or “inferred” (generated from the title and metadata only), so you can judge how much weight an entry deserves.

Several content types on CultCodex are generated by AI (Anthropic Claude) based on episode transcripts and metadata. These include:

• ·Episode summaries (short and long)
• ·Guest / character profiles ("Codex entries")
• ·Notable quotes extraction
• ·Lore entries and topic clusters
• ·Psychenomicon narrative chapters (explicitly mythic/symbolic interpretations)

AI-generated content is clearly marked with a ◈ AI-generated notice wherever it appears.

• ·Summaries use neutral, descriptive language without editorial judgment.
• ·Guest identifications are based on in-stream introductions and display names.
• ·Quotes are extracted verbatim from transcripts where possible.
• ·Topic tags are normalized to avoid duplicates (e.g. “AI” vs “Artificial Intelligence”).
• ·Lore entries distinguish canonical (stated on stream), speculative, and community myth.
• ·Person pages carry an archive-context notice explaining that profiles are auto-generated.

AI summaries and character profiles describe on-stream discussion and performance personas — not verified real-world claims.

• ·Summaries are generated from transcripts and may misattribute statements, misidentify speakers, or miss context.
• ·Character profiles reflect how someone appears across recorded streams, not their private life.
• ·AI models can hallucinate — descriptions may contain factual errors.
• ·Transcript accuracy varies: auto-generated captions and Whisper both mishear words and names, especially in overlapping speech or low-quality audio.
• ·Some early episodes have no transcript at all and only minimal metadata.
• ·Guest identification relies on display names, which may not reflect legal or preferred names.
• ·Topic and lore categorization involves subjective judgment by the AI model.
• ·Psychenomicon chapters are explicitly symbolic/mythic interpretations, not factual reporting.

If you find inaccurate, harmful, or outdated information, please use the correction form linked on every episode and person page.

Typical turnaround for corrections is 1–3 days. For urgent removal requests (e.g. content that puts someone at risk), contact the site owner directly.

• Video metadataYouTube Data API v3
• TranscriptsYouTube auto-generated captions via Supadata API; OpenAI Whisper (medium) run locally when captions are unavailable
• AI enrichmentAnthropic Claude (claude-opus-4-8)
• Semantic searchOpenAI text-embedding-3-small + pgvector
• StoragePostgreSQL on Xata, via Prisma