Random Thoughts

Personal project, January–February 2026. Site offline.

A real-time simulation of a thousand people who live, form relationships, age and die. An LLM writes their inner thoughts into a shared feed. A prompt engineering project.

The live feed: short first-person thoughts from a 13-year-old, a 25-year-old, people in their forties and fifties, and a 90-year-old, each tagged good or evil and as an everyday thought or a birthday — Shared feed.

Simulation

People. Name, age, cultural background, Big Five traits, life themes, relationships and a moral score from −100 to +100.
Time. A minute is a day: birthdays, births, marriages, divorces, deaths. The dead are reborn with karmic weight.
Thoughts. Written from each person's circumstances in an age-appropriate voice, tagged good, neutral or evil.

Prompt Engineering

Each request stacks three layers:

Segments. Global rules, toggled at runtime: no narration, personality guidance, time of day, season, cultural context, output format, moral classification.
Template. Per thought type (everyday, birthday, dying, whisper), with voice guidance per age band.
Context. Built from the person's state: date, season, age, heritage, Big Five traits, obsessions, moral compass, relationships.

Big Five traits become plain phrases in the context. A segment maps each phrase to thought style.

Trait	High	Low
Openness	creative and curious	conventional and practical
Conscientiousness	disciplined	spontaneous
Extraversion	outgoing	reserved
Agreeableness	cooperative	competitive
Neuroticism	emotionally sensitive	emotionally stable

Same night, opposite personalities

Context

Current date: November 3, Year 12 at 11:40 PM (autumn's change, a time of reflection and letting go)
Time of day: night
Person: @318, Age: 34
Cultural heritage: Hispanic
Personality: creative and curious, spontaneous, outgoing, cooperative, emotionally sensitive
Obsessions: creativity, relationships, money
Moral compass: 40% good, 0% evil
People in their life: spouse: @402 [82%], friend: @77 [64%], coworker: @921 [31%]

{"thought":"Can't sleep. Sketching @402 snoring and it's the best thing I've drawn all month. Rent's due Friday and I spent the grocery money on oil paints, but they were the exact orange of the leaves outside.","moral":"good"}

Context

Current date: November 3, Year 12 at 11:40 PM (autumn's change, a time of reflection and letting go)
Time of day: night
Person: @921, Age: 34
Cultural heritage: Asian (Korean)
Personality: conventional and practical, disciplined, reserved, competitive, emotionally stable
Obsessions: work, money, legacy
Moral compass: 0% good, 18% evil
People in their life: parent: @210 [71%], sibling: @664 [48%], coworker: @318 [31%]

{"thought":"Deck finished, backed up twice, alarm set for 5:10. If @318 charms the client with my numbers again, the VP gets the version history.","moral":"evil"}

Whispers

A 90-year-old's page: Big Five personality, a moral score of +7 with a killer tendency of 56%, the Divine Whisper box, and her feed showing the whisper, her reply, and an earlier thought

Whisper. Visitors can whisper to anyone. The model judges it, adjusts that person's moral score, and writes a reply in their voice.
Example. A 90-year-old writing the family history at 2 a.m. is asked which story the family should remember. Judged good: score −9 to +7. Reply: the winter the family lost the farm.
Limits. Cooldown and daily cap per recipient.

Relationships

Bonds drift daily. Acquaintances above 60% can become friends, friendships below 20% can end, ex-friends above 70% can reconcile. Family ties are fixed.

The population's moral balance bar with a tooltip on the evil segment: 0 Very Evil, 97 Evil, 182 Mostly Evil — Population moral balance. Hover for a breakdown by tier.

Systems

LLM queue. Requests spread across API keys and providers (Groq first, Google fallback), with rate limiting, per-key health and a daily budget cap.
Admin. Server metrics, logs, security audit. World state persists across restarts.

The epiphany dashboard with toggles for each thought type and each prompt segment, and the prompt template