LinkedIn Personal Brand Agent

To solve the authenticity and specificity trade-off, we engineered the LinkedIn Personal Brand Agent—a Next.js and Supabase-powered platform featuring a Three-Layer RAG Architecture and an iterative Writer-Critic Self-Correction Loop.

graph TD
    A[Raw Developer Note] --> B[Auto-Structure RAG Pipeline]
    B --> C[(Layer 1: Raw Post Intelligence)]
    B --> D[(Layer 2: Blueprint Libraries)]
    C & D --> E[Parallel Vector Query RPCs]
    E --> F[Writer Agent: Draft Generation]
    F --> G[Critic Agent: Editorial Audit]
    G -- "Fails Score < 85 or Cliches Found" --> H[Self-Correction Loop + Feedback]
    H --> F
    G -- "Passes" --> I[Approval Deck & Publish Queue]

1. The Three-Layer Content Intelligence

Rather than sending flat prompts to the LLM, the system dynamically curates a highly specific prompt context by querying distinct database vaults:

• Layer 1: Raw Post Intelligence (raw_post_intelligence): An append-only database housing unedited, high-performing posts. The generation engine retrieves Success Cases (posts with hook strengths >= 8) to mimic positive structures and Failure Cases (posts with low specificity or high cliché density) to explicitly avoid.
• Layer 2: Copywriting Blueprint Libraries: Four separate libraries storing compiled, deduplicated templates:
  • Hooks Library: Structures proven to hook attention under 140 characters.
  • Story Frameworks: Narrative blueprints (e.g., Problem $\rightarrow$ Attempt $\rightarrow$ Failure $\rightarrow$ Lesson).
  • CTA Library: Direct, context-driven audience engagement prompts.
  • Anti-Patterns: Explicit blacklists of corporate buzzwords and AI jargon.
• Layer 3: Dynamic Rankings & Semantic Search: Using Google's gemini-embedding-001 model, the agent generates 3072-dimensional vector embeddings for incoming topics and user inputs. It performs nearest-neighbor search via Supabase custom SQL functions (RPCs) using cosine distance: $\text{Similarity} = 1 - (\vec{A} \cdot \vec{B})$

2. The Writer-Critic Self-Correction Loop

The system coordinates two specialized LLM instances to refine drafts before they reach the user:

• The Writer Agent: Generates drafts using retrieved memory logs, convictions, blueprints, and style guidelines. It is instructed to write in an honest, vulnerability-first tone, incorporating casual punctuation and developer-specific details.
• The Critic Agent (Temperature 0.1): Acts as a ruthless chief editor. It audits the draft against a strict scoring rubric (overall score out of 100). Any draft containing blacklisted words or scoring below 85/100 is automatically rejected, returning actionable feedback to the Writer for regeneration.

3. Responsive, Tabbed UX Design

To make the complex platform operational on the go, we refactored the interface layout:

• Desktop Layout: Displays the full command center (Memory Vault, Creative Studio, Approval Deck, and Audit Suite) side-by-side in a responsive grid.
• Mobile Layout: Implements a segmented tab controller that collapses the workspace. Users focus on one phase at a time (e.g., entering raw notes, executing generation, or reviewing drafts) with clean, tailored inputs, cards, and warm ivory/orange aesthetics.