Grounded identity
Make agents feel like they belong in the room.
Persona6 turns conversation history into grounded persona graphs. Start with Discord communities, recover the member types that already shape the room, then use those graphs to keep agents useful, native, and inspectable.
Community thread
“I always notice when a product forgets my priorities between sessions.”
- continuity-sensitive
- high recall
- strategy-first
Persona graph
P6
memory
character
skill
Why this persona fits
- stable preference stack across sessions
- repeatable language texture under new prompts
- consistent task strategy and failure tolerance
Use case
Start with one server. Prove identity there.
The wedge is narrow on purpose. Discord gives rich behavioral evidence, obvious failure cases, and a room where continuity matters immediately. That makes it the right place to prove grounded personas before broader product surfaces.
Discord wedge
One room, five recurring member types, one inspectable runtime.
Persona6 should read the room the way an operator does. Not just what people said once, but how they repeat preferences, how they switch tone, what they tolerate, and how they ask for help.
- Read channel history and recurring interaction patterns.
- Cluster member archetypes without flattening the community.
- Build persona graphs across memory, character, and skill.
- Use those graphs to constrain agent replies in live threads.
- Measure whether conversations stay useful and native to the room.
Fresh-chat recognition
5 candidates / 1 grounded matchThis is the cleaner proof surface than a generic chatbot demo. If the system can recognize the right persona from a fresh exchange and explain why, the rest of the runtime becomes easier to trust.
System
Three persona layers, plus one operator layer.
Identity needs structure. Persona6 separates what someone remembers, how they tend to act, what they can reliably do, and what an operator should still control in the loop.
Proof
The product has to survive contact with a live room.
The real test is not whether the copy sounds smart. It is whether the runtime preserves continuity, stays useful under pressure, and gives operators enough evidence to trust what it is doing.
Runtime checks
What we should measure in the first pilot
- Reply usefulness versus baseline human or bot responses.
- Correction rate after agent interventions.
- Thread depth after an agent reply lands.
- Persona stability across fresh prompts and new sessions.
- Operator trust in the evidence shown beside each decision.
Operator scoreboard
first pilot targets
Grounded match rate
90%+
Correction rate
<10%
Receipt coverage
100%
Operator review lag
<2 min
The UI should make these targets obvious. If a persona graph cannot show receipts or a match falls below confidence, the operator should see that immediately instead of reverse-engineering it after the fact.
Blog
Write around the work, not around a content calendar.
The blog should feel like an extension of the product surface: clear research notes, concrete runtime lessons, and articles that are easy to update because they share the same visual system as the landing page.
Discord use case
How to put grounded agents inside a Discord server
Why Discord is the right first surface and how to measure whether agent contributions are actually helping the room.
Research lens
What Anthropic's persona selection model means for product builders
A useful mental model for assistant personas, and why the user side needs just as much structure.
Foundational idea
Memory Is Not Identity
Why storing facts from a user is not enough to preserve the person across sessions.
FAQ
Short answers for operators and builders.
Is this a moderation bot?
No. Persona6 is a grounded persona layer first. Moderation can be part of the operator workflow, but the product focus is continuity and native-feeling conversation quality.
How easy is it to rebrand later?
The shared CSS already runs on color tokens in one place. Landing and blog now use the same style system, so brand updates should be much cheaper.
Why keep the public site simple?
The homepage should explain the product thesis, the first wedge, and the proof surface. Deeper workflow detail can live in the blog and the workbench without turning the landing into a dump of internal thinking.
Next step
If you want agents in a live room, start with one server.
Best fit right now: teams running an active Discord community who want better continuity, less generic automation, and a persona system they can inspect.
Links
Talk, read, or inspect the repo.
The public surface should stay simple: one booking path, one blog path, one GitHub path.