Prompt Index

This service only ships prompts that are used by active flows or media processors. Executable user prompt templates and instantiated runtime requests are documented in Runtime Prompt Requests.

Active Flow Prompts

All active flow prompts return raw JSON objects in Spanish. The model must not wrap the response in Markdown or a ```json block. Markdown is allowed only inside JSON string values when a parent-facing report benefits from formatting. Report and memory prompts also treat missing/untrusted clocks, technical failures and status/channel metadata as non-evidence: they cannot create risk, chronology or an absence-of-activity conclusion.

Prompt	Used by step	System content	User content
prompts/smoke_test.md	smoke_test	Runtime health-check rules and required JSON shape.	Rendered conversation documents.
prompts/initial_report.md	initial_report	Parent report context, anti-false-positive rules, tutor guidance and age-band boundaries, group/shared-group context, risk taxonomy, name rules, JSON shape and report.markdown.	Monitored child JSON, generic context without memories/guidance/age band, optional normalized tutor guidance and known age band, optional user stats and rendered conversation documents.
prompts/chat_id_summarization.md	chat_id_summarization	Relationship inference, group/shared-group context, compact memory initialization/update decisions, first-name rules, cross-chat context, required chat_name, chat_id_summaries, and chat_id_memory_updates JSON shape. Runtime may use internal batches and retry missing chat ids without changing the public output contract.	Runtime template: monitored child minimal identity JSON (name) and rendered conversation documents with matching per-chat memories for the current internal batch.
prompts/daily_summary_and_notifications.md	daily_summary_and_notifications	Daily summary, notification vocabulary, taxonomy, tutor guidance and age-band boundaries, stats and memory-update rules.	Runtime template: Markdown sections for monitored child, run metadata, stats, optional tutor guidance and known age band, prior reports and day documents with per-chat memories. Raw side-contract fields are excluded from generic context and memory decisions.
prompts/interweek_report.md	interweek_report	InterWeek synthesis, taxonomy semaphore, trajectory, tutor guidance and age-band boundaries, stats context, parent action and Markdown rules.	Runtime template: Markdown sections for monitored child, run metadata, stats, optional tutor guidance and known age band, and prior reports. No conversation documents or per-chat memories in v1.
prompts/content_trust_evaluator.md	content_trust_evaluator	Independent instruction-boundary, evidence and memory grounding checks with a strict typed response.	Canonical untrusted envelope containing final typed output and derived conversation documents.

All active flow prompts are called through services.model_gateway.OpenRouterJsonModelClient. Primary model is google/gemini-3.5-flash; default shadow model is google/gemini-3.1-flash-lite. The primary output is authoritative. Shadow output is stored on the run result and traced in Langfuse when enabled.

Active Media Prompts

Prompt	Used by activity	System content	User content
prompts/vision.md	analyze_image, video frames, embedded document images	Spanish visual description rules, first-name guidance for descriptions, literal visible-text transcription, and required JSON shape.	Image content plus optional caller context.
prompts/video.md	process_video for video and animated sticker/GIF content	Whole-video visual/audio analysis rules and the structured video output contract.	Complete video content, including locally converted animated media.

Audio transcription does not use a prompt file.

Trial Prompt Sets

Trial prompt sets are trial-local copies used for prompt iteration, not active runtime prompts. They live under the trial folder's prompt_sets/<prompt-set>/prompts/ directory with a prompt_set_manifest.json. Replays snapshot the effective files into replays/<prompt-set>/00_prompts_used/, and replay model calls use that snapshot.

Use prompt sets when comparing prompt-v1, prompt-v2, or prompt-v3 outputs from an existing trial without re-running media processors. Keep active prompt files in this directory reserved for repo-level runtime changes.

Prompt Rules

• Put stable rules, output contracts, and compliance constraints in System.
• Put variable data only in User: rendered conversations, monitored child JSON, stats metadata, image/audio content, or batch chat documents.
• Flow, vision, and video outputs must be valid raw JSON objects in Spanish.
• Markdown may appear inside JSON strings only; never wrap the whole response in Markdown.
• Use only first names for the monitored minor and named contacts when available, except literal transcription fields such as texto_visible.
• Do not list prompts here unless a recipe, activity, or model gateway uses them.
• Do not build prompts from raw upstream payloads; adapters and the conversation builder run first.
• Treat [MEDIA_NO_PROCESADA] as a technical evidence-boundary marker only;

parent-facing outputs must not mention missing or unprocessed media.