Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/LilMGenius/polysona/llms.txt

Use this file to discover all available pages before exploring further.

Polysona is built around five specialized agents that work in sequence to transform deep psychological self-knowledge into published, persona-authentic content. Rather than relying on a single generalist model, Polysona separates extraction, trend detection, drafting, quality assurance, and publishing into distinct agents — each with a focused role, a dedicated toolset, and strict output contracts. The pipeline is portable: every agent runs identically in OpenAI Codex (via $command) or Claude Code (via /command).

​
Agent Summary

AgentRoleCodex CommandClaude Code CommandBadge
profilerDeep psychology interviewer$interview/interview10 frameworks
trendsetterTrend detector$trend/trendWebSearch
content-writerPlatform content generator$content [platform]/content [platform]5 platforms
virtual-followerQA simulator$qa/qacontext: fork
adminPublisher and tracker$publish/publishfeedback loop

​
Pipeline Phases

The Polysona pipeline is divided into two distinct phases: SETUP and LOOP.
1

SETUP Phase — Run once (or periodically)

The profiler conducts a structured interview using 10 psychology frameworks. After the interview, Polysona’s structuring engine processes the raw interview log and produces three persona files:
  • personas/{id}/persona.md — core identity, decision patterns, energy model, blind spots
  • personas/{id}/nuance.md — voice register, platform style, phrasing constraints
  • personas/{id}/accounts.md — rolemodel accounts and virtual follower benchmarks
This phase is run once for a new persona and repeated whenever a meaningful identity shift is detected.
2

LOOP Phase — Runs every content cycle

With persona data in place, the four remaining agents form a repeating loop:
  1. trendsetter → scans for ranked trending topics matched to persona domains
  2. content-writer → generates 3 platform-native draft variations per selected topic
  3. virtual-follower → simulates 20 audience archetypes and returns a scored TOP 5
  4. admin → publishes the selected draft, records metadata, and feeds engagement results back into the persona
You drive the loop: the user selects a draft from the virtual-follower’s TOP 5 before admin publishes it.

​
Pipeline Architecture

┌─────────────────────────────────────────────────────────────┐
│                         POLYSONA                            │
│      Build and run multiple personas across any AI agent.   │
│                                                             │
│  SETUP (one-time / periodic refresh)                        │
│    ① profiler -> interview log extraction                   │
│           -> Polysona structuring engine                    │
│           -> persona.md + nuance.md + accounts.md           │
│                                                             │
│  LOOP (per content cycle)                                   │
│    ② trendsetter -> ③ content-writer -> ④ virtual-follower │
│    -> user selection -> ⑤ admin (publish + tracking)       │
└─────────────────────────────────────────────────────────────┘

​
Agent Invocation

Polysona agents are invoked by command — the syntax differs slightly between the two supported AI agent hosts.
In OpenAI Codex, AGENTS.md is auto-recognized and skills are auto-discovered from .agents/skills. Use the $ prefix to invoke any agent command:
$interview        # Run the profiler interview
$trend            # Scan trending topics
$content x        # Generate X (Twitter) drafts
$qa               # Run QA simulation
$publish          # Publish selected draft
If you edit skills/, resync with:
node ./scripts/sync-codex-skills.mjs
# or
bun run codex:skills:sync
Utility commands $introduce / /introduce, $status / /status, and $export / /export are also available across both hosts for session injection, pipeline status checks, and persona export.

​
The context: fork Isolation Model

The virtual-follower agent runs with context: fork, which means it executes in a context isolated from the content-writer’s generation context. This is intentional: the QA simulation must evaluate drafts as an independent reader, not as a continuation of the writing session. Without this isolation, the evaluator would carry forward the same assumptions and blind spots as the generator — defeating the purpose of multi-audience review.
Do not manually pass context from a content-writer session into a virtual-follower session. The context: fork boundary exists to guarantee evaluation independence.

​
Agent Pages

Profiler

Deep psychology interview engine. Runs a structured 50-turn session using 10 frameworks to extract unconscious patterns and build your three core persona files.

Trendsetter

Persona-matched trend detector. Scans domains and returns a ranked list of 5 topics with platform-fit tags conditioned on your active persona.

Content-Writer

Platform-native draft generator. Produces 3 persona-conditioned variations per platform, adapted to the voice and hook patterns in your nuance.md.

Virtual-Follower

QA simulator with 20 audience archetypes. Scores drafts across hook, empathy, shareability, CTA, and platform fit — then recommends a TOP 5.

Admin

Publishing flow and performance tracker. Saves final content, generates a platform-specific checklist, and captures metadata for the feedback loop.

Build docs developers (and LLMs) love

Get started for freeTalk to us