How AI Content Generation Works

Not Another GPT Wrapper

Most AI writing tools work the same way: you type a prompt, it goes to GPT, you get generic output. The AI has no idea who you are, what you've read, or what you actually think.

Postiller is different. When you generate a post, the AI draws from:

• Bookmarks you've saved and annotated
• Ideas you've captured
• Notes explaining why content matters to you
• Your defined persona and voice

The result isn't generic AI slop. It's content grounded in your specific knowledge and perspective.

The Problem with Generic AI

Ask ChatGPT to write a LinkedIn post about leadership. You'll get something like:

"Leadership isn't about having all the answers—it's about asking the right questions. Great leaders empower their teams, foster collaboration, and lead by example. What leadership lessons have shaped your journey? #Leadership #Growth #Teamwork"

It's not wrong. It's just... nothing. Generic advice wrapped in generic formatting with generic hashtags.

The AI doesn't know:

• What articles about leadership you've actually read
• Which ideas resonated with you and why
• Your specific experiences or expertise
• How you actually talk

So it produces content that could have been written by anyone—which means it sounds like no one.

Retrieval-Augmented Generation (RAG)

RAG is the technique that makes Postiller different. Instead of generating from nothing, we first retrieve relevant context from your personal knowledge base.

How It Works

Step 1: Understand what you're writing about

When you start generating a post—whether from a bookmark, an idea, or a blank slate—Postiller identifies the core topics and concepts.

Step 2: Search your knowledge base

Using semantic search, we find content in your collection that's conceptually related:

• Bookmarks about similar topics
• Ideas that connect to the theme
• Notes you've written about why things matter

Step 3: Build context for the AI

The relevant content gets assembled into a context package. This includes:

• Extracted article content (key passages, not full text)
• Your notes and learnings
• Your idea descriptions
• Your persona guidelines

Step 4: Generate with context

The AI receives this context along with your request. It's not starting from its general training—it's starting from your specific knowledge.

Your Notes Matter Most

Here's something most AI tools miss: your annotations are more valuable than the content you save.

When you add a note to a bookmark saying "Great framework for thinking about team scaling challenges," that note captures why the content matters to you. The original article might be 2,000 words about various topics—your note distills what's actually relevant.

Postiller's RAG system prioritizes user-generated content:

This weighting ensures the AI understands your perspective, not just the raw material you've collected.

Context Assembly

Not all relevant content fits in an AI prompt. We use chunking and selection to build the most useful context:

Smart Chunking

Long articles are split into ~400 character segments, each capturing a single idea. When your post topic matches a specific section of an article, we include that section—not the entire piece.

Relevance Scoring

Each chunk is scored for relevance to your current task:

• Semantic similarity to your topic
• Recency (newer content weighted slightly higher)
• User engagement (content you've annotated scores higher)

Context Window Management

AI models have limits on how much text they can process. We select the highest-scoring chunks that fit within these limits, ensuring the AI gets the most relevant context possible.

Persona Integration

Context is half the equation. The other half is voice.

When you generate content, your active persona shapes how the AI writes:

Communication style — Direct or conversational? Formal or casual? Your persona defines the tone.

Expertise areas — What topics do you know deeply? The AI emphasizes your areas of authority.

Phrases to use — Language patterns that sound like you.

Phrases to avoid — Clichés and patterns that sound generic (goodbye "excited to announce").

The AI receives both your relevant content AND your persona guidelines. The result is informed by your knowledge and sounds like your voice.

The Generation Flow

Here's what happens when you tap "Generate" on a bookmark:

1. Extract topic from bookmark
   └── "Article about async communication in remote teams"

2. Semantic search your knowledge base
   └── Find 12 related chunks from 5 bookmarks
   └── Find 3 related ideas
   └── Include your notes from the source bookmark

3. Score and select context
   └── Pick top 8 chunks by relevance
   └── Total context: ~2,500 tokens

4. Load active persona
   └── "Professional Voice: Direct, thoughtful,
        expertise in team building"

5. Construct prompt
   └── System: Persona guidelines
   └── Context: Selected knowledge chunks
   └── Task: Generate LinkedIn post

6. Call AI provider (your API key)
   └── Response: Draft post

7. Present for editing
   └── You refine and publish

The whole process takes 3-5 seconds. Most of that is waiting for the AI provider to respond.

Why This Produces Better Content

Generic AI writing fails because it lacks two things: knowledge and voice.

Knowledge gap: ChatGPT knows a lot about leadership in general. It doesn't know that you recently read an article connecting military decision-making frameworks to startup pivots, and you found the "OODA loop" concept particularly useful for your current situation.

Voice gap: ChatGPT can write in a "professional" tone. It can't write in YOUR tone—the specific way you structure arguments, the analogies you prefer, the phrases that feel natural to you.

RAG plus personas close both gaps:

• Your knowledge: The AI references specific articles and ideas you've collected
• Your voice: The AI follows your persona guidelines

The content still needs editing—AI isn't perfect. But you're starting from something informed and authentic, not generic filler.

Practical Examples

From Bookmark to Post

Saved bookmark: Article about "The hidden costs of context switching"

Your note: "Good data on the 23-minute recovery time. Connects to our standup restructuring."

Generated post:

The research is clear: every interruption costs 23 minutes of recovery time. Not because the interruption itself takes that long—but because your brain needs time to rebuild context.

We restructured our standups around this insight. Instead of synchronous morning calls that fragment everyone's focus, we moved to async updates with a single "focus hours" block where Slack goes quiet.

Result: deeper work, fewer "quick questions" that aren't quick, and a team that actually finishes things.

Notice how the post:

• References specific data from the article (23 minutes)
• Connects to your note about standup restructuring
• Draws from your apparent experience with team processes
• Avoids generic "leadership lesson" framing

From Idea to Post

Your idea: "Thread idea: Why most productivity advice fails for creative work"

Related bookmarks found:

• Article on maker vs manager schedules
• Article on creative incubation time
• Your note: "Productivity gurus optimize for output, not insight"

Generated post:

Productivity advice is written by managers for managers. Calendar blocking. Task batching. The two-minute rule.

Creative work doesn't operate on those terms. Ideas need incubation. Insights emerge from what looks like "doing nothing." The most productive creative session might be a walk where something finally clicks.

The maker schedule exists because creative work has different physics than administrative work. Optimizing for busyness kills the conditions that produce actual breakthroughs.

The AI connected your idea to relevant bookmarks you'd saved, producing a post that synthesizes multiple sources into a coherent argument.

What RAG Can't Do

Let's be clear about limitations:

Can't create expertise you don't have. If you haven't saved content about a topic, the AI has nothing to work with. RAG retrieves from your knowledge base—it doesn't fabricate credentials.

Can't replace editing. Generated content is a strong first draft, not a final product. You should always review, refine, and ensure it accurately represents your views.

Can't guarantee accuracy. The AI might misinterpret your notes or make connections you didn't intend. You're responsible for fact-checking before publishing.

Can't make boring topics interesting. If your source material is dry, the output will be too. Garbage in, garbage out—just with better grammar.

RAG makes AI writing useful by grounding it in your knowledge. It doesn't make AI writing perfect.

Building Your Knowledge Base

The quality of generated content depends on what you've collected. A few tips:

Save intentionally. Don't bookmark everything—bookmark what resonates. Quality over quantity.

Add notes. A two-sentence note about why you saved something is worth more than the entire article for generation purposes.

Capture ideas. Even half-formed thoughts give the AI material to work with. "Something about how trust compounds" is better than nothing.

Review and refine. Periodically look at what you've saved. Delete irrelevant stuff. Add notes to things that matter.

Your knowledge base is the fuel. The better the fuel, the better the output.