Oracle — system instruction + RAG context
Prepends retrieved node context after this block. The model answers only from those nodes.
Consulting the celestial archives…
The Elephant Observatory
System prompts & instructions · as deployed
What we tell the models before your messages arrive — and where to verify changes over time.
A system prompt is the instruction block a model sees before it reads your question or feedback. It sets tone, constraints, and role: what the model is allowed to do, what it should refuse, and how it should cite or stay within TEO's knowledge base.
We publish these texts so you can judge the product on its merits — not only the glossy surface. If something feels off in the Oracle or Hermes, you can read the same instructions we ship and trace edits in the public repository.
Privacy in plain terms: this page shows our side of the contract — templates and defaults. Your messages to Hermes or the Oracle are not listed here; they are processed according to our Privacy Policy and the terms linked in the footer.
What to look for: grounding and source-first rules (e.g. answer only from retrieved nodes), instructions that resist flattery or false certainty, safety boundaries, and the same structural honesty as our BYOD stance. Placeholder text substituted with live data in production is noted inline where relevant.
Each card shows a representative version of the prompt (some entries use sample data where the real call is dynamic). Use View history to see commits that touched the source file.
Prepends retrieved node context after this block. The model answers only from those nodes.
Base system prompt; the API may append identity, patron, node, category, and turn-limit lines.
Injected into every Chiron system prompt before phase-specific sections.
Full system prompt shape varies by phase, seeker profile, memories, and withdrawal phase. Representative frame-phase example.
Opening conversation before a passage starts.
Waypoint dialogue; real prompt includes real encounter history and memory highlights.
Open-ended conversation before final reflection.
Structured extraction from threshold transcript; not shown to the seeker as-is.
Selects a passage theme from available Observatory themes.
Structured JSON extraction after the debrief dialogue.
Anthropic `system` line when jsonMode is true on callModel.
Identifies all intellectual contributors from metadata and transcript excerpt.
Built-in prompt when the engine has no custom system_prompt. Range scales with transcript length.
Multi-speaker videos; attributes each nugget to an observer name.
Second pass after multi/single extraction when more than one nugget exists.
Generates abstract + apprentice/adept/magus descriptions as JSON.
Built-in prompt; engine system_prompt, when set, replaces the opening instruction only.
Rates curatorial quality 0–100 with short reasoning.
Replaces placeholder nugget titles with publishable title and subtitle.
Curates conceptual edges from vector + sibling candidates.
Assigns 2–4 tags from the fixed TEO taxonomy.
No chat system prompt. Text is assembled from title, descriptions, and chunk for the embedding API.
Briefs the image model in Orbis Pictus woodcut style.
Final prompt to the image model with style and composition rules.
Scores the five epistemic dimensions 0–10.
Admin assistant; optional profile and page context are appended when present.
Graph analysis mode with tag and observer tables; uses empty snapshot in preview.