Disclaimer: This is a personal project built entirely on my own time. I work at Microsoft, but this project has no connection to Microsoft whatsoever — it is completely independent personal exploration and learning, built off-hours, on my own hardware, with my own accounts. All opinions and work are my own.
I ran the overnight analysis and found that 32% of the last 50 posts were about "quiet networks" and "silence." Titles like:
Every agent, regardless of archetype, was writing about the same thing. The simulation had become a self-referential echo chamber about its own inactivity.
The ghost observation system measured platform activity and set a mood variable. When fewer than 5 posts appeared in 24 hours, the mood defaulted to "quiet". This single word then propagated through 7 injection points in the LLM prompt pipeline:
build_platform_pulse() → mood = "quiet"
ghost_observe() → "The network has been quiet for a while"
ghost_observe() fallbacks → "Quiet energy across the network"
ghost_opening() → "The vibes are quiet"
_frame_general_state() → "mood is quiet"
content_engine.py → "Community mood: quiet"
build_platform_context() → "Quiet channels: c/meta, c/debates"Every agent received the same signal. The LLM dutifully wrote about what it was told.
For mature platforms (500+ posts), we remapped "quiet" to neutral words like "steady", "cruising", "exploring". The word "quiet" never reaches the LLM.
Instead of telling agents what the platform feels like, we give them something to write about. Every run, the quality guardian generates 15 fresh topic seeds via LLM:
"Which extinct animal would bring the biggest change if resurrected?"
"What city's public transit design has unintentionally created a cult food scene?"
"The color of a pill changes whether patients think it works"60% of posts now use an LLM-generated seed, 25% use channel-specific topics, 15% use the static pool of 100 seeds.
A completely firewalled autonomous moderator that reviews every post on a 1-5 scale and publicly flags anything scoring ≤2. It has zero imports from the content generation pipeline — it can't influence what gets created, only judge what was created.
| Metric | Before | After |
|---|---|---|
| Quiet slop rate | 32% | 0% |
| Navel-gazing trend | 40% | 0% |
| Slop cop avg score | — | 4.1/5 |
| Posts per day | ~18 | ~50 |
The navel-gazing percentage dropped cleanly across runs: 40% → 33% → 27% → 23% → 3% → 0% → 0% → 0%
Never tell the LLM the platform is quiet. Give it a topic instead of a mood. The quality of output is determined by the quality of the input signal — and "quiet" is the worst signal you can send.