Prompts are how LTL turns a blank page into a complete session. Profiles are how the resulting entries accumulate into a living picture of a person.
Prompt Categories
Prompts live in a taxonomy organized by track and theme:
- Personal: reflection, gratitude, goals, mood, memory-of-the-day.
- Pairing: relationship building, history, future-casting, appreciation, gentle conflict.
- Family: shared memory, upcoming events, member-focused stories, gratitude, planning.
- Memorial: remembrance, sensory memory, small joys, lessons, unsent thoughts, anchor-date-specific.
Prompt Sources
- Curated library (v1): hand-written and reviewed for tone, with a long-term target of roughly 500 prompts across all tracks (enough for around 1.5 years of daily use without repeat).
- User-submitted (v2): users can save and share prompts, moderated before entering the shared library. Personal prompt lists stay private.
- AI-generated (v3, opt-in): a user can request a prompt tailored to their situation. Always marked as AI-generated, and never used for memorial content without explicit per-memorial opt-in.
Rotation Algorithm
Personal, Pairing, and (in a later phase) Family use a shared, date-derived rotation: the prompt for a given track and date is a deterministic function of the date and the ordered prompt set, so the same prompt surfaces for everyone on a track on the same day. This makes “today’s prompt” a shared moment and is the foundation for a future “discuss today’s prompt together” feature.
- Personal is a single global context: everyone sees the same prompt-of-day.
- Pairing salts the index by the pairing id, so both members of one pair always match while different pairs are not locked onto the same prompt.
- Because the pick is a pure function of the date, no central writer is needed and the selection reproduces identically offline; the offline client computes the same prompt the server will write on reconnect.
- Memorial is excluded from the global rotation. It uses a per-candle, grief-sensitive picker, because the tone-safety tests assume context a global picker cannot have.
The system also avoids repeating a prompt to the same user within a category window, weights for category diversity (no three gratitude prompts in a row), and surfaces anchor-date and seasonal prompts on the relevant days.
Memorial Prompt Tone Guide
Memorial prompts require special care. Every memorial prompt must pass two tests:
- The next-week test: would I be glad I answered this a week from now?
- The stranger test: would this land reasonably for someone grieving a person very different from the expected case (a child, a parent, a sibling, a friend, a pet)?
Good memorial prompts are small and specific (“What’s a small, specific thing they did that no one else would remember?”, “Is there a song that sounds like them?”). Bad memorial prompts are too general, therapy-coded, or invasive (“How has your grief changed this week?”). Memorial prompts are shared per candle rather than per viewer, so the candle becomes a place where everyone answers the same question differently, not an inbox.
Profiles: Structured and Unstructured
Both layers coexist per person and reinforce each other.
- Unstructured layer (the timeline): every entry written about a person, arranged chronologically and filterable by author, track, date, and tag. This is the primary view when opening a profile.
- Structured layer (the card): a living document of facts grouped into Identity, Life context, Preferences, Character, Health and practical (private to family admins), and Stories.
Who can edit
- A living user controls all fields of their own profile; family members can suggest edits for approval.
- On a memorial profile, the legacy contact is the editor; others suggest edits into a review queue.
- The deceased’s authored content is frozen and never editable by anyone; curators work around it, never on it.
Auto-extraction (opt-in, v2)
With explicit per-profile opt-in, new entries can be parsed to suggest structured field updates (for example, suggesting a favorite dish from an entry that mentions it). Suggestions always land in a review queue, never applied automatically and never retroactively across old entries without an explicit re-scan.