Why Three-Pass AI Summarization Beats Single-Shot

When we first built Huddix's summarization system, we started with a simple approach: transcribe the meeting, feed the transcript to an LLM with a prompt like "summarize this meeting," and return the result. It worked — but the summaries were often generic, missed key points, and lacked the nuance that makes meeting summaries truly useful.

After extensive experimentation, we developed a three-pass pipeline that produces dramatically better results. Here's why.

The Problem with Single-Pass Summarization

Single-pass summarization has several limitations:

• Lack of focus — A single prompt must balance multiple objectives (key decisions, action items, context, next steps) simultaneously, often doing none perfectly.
• Inconsistent structure — Without explicit guidance, LLMs produce summaries with varying structures that are hard to compare across meetings.
• Missing nuances — Important details get lost in the generalization process, especially when the prompt doesn't explicitly ask for them.
• Factual errors — LLMs can hallucinate details when summarizing long transcripts in a single pass.

Our Three-Pass Pipeline

Pass 1: Structured Extraction

In the first pass, we use Claude Haiku to extract structured data from the transcript: speakers identified, topics discussed, questions raised, and any explicit commitments made. We use a JSON schema to ensure consistent structure and minimize hallucination.

Pass 2: Thematic Analysis

The second pass analyzes the extracted data to identify themes and patterns. Which topics received the most discussion? Where did people disagree? What was the overall sentiment? This pass produces a high-level narrative that captures the meeting's context.

Pass 3: Summary Generation

The final pass synthesizes the structured extraction and thematic analysis into a coherent, well-structured summary. We explicitly prompt for key decisions made, action items with owners, questions to follow up on, and a brief narrative overview.

Real Comparisons

We tested both approaches on 100 meetings and had human evaluators rate the quality. Results:

• Completeness: 4.2/5 (3-pass) vs 2.8/5 (single-pass)
• Accuracy: 4.5/5 vs 3.9/5
• Action item extraction: 92% vs 67%
• Useful for later reference: 4.4/5 vs 2.9/5

Why Claude Haiku?

We evaluated multiple models for this pipeline. Claude Haiku offered the best balance of speed, cost, and quality for our use case. Its instruction-following capabilities are particularly strong for structured extraction tasks, and it's fast enough to process meetings in real-time.

Conclusion

The three-pass pipeline isn't just about more processing — it's about decomposing a complex task into simpler subtasks that can each be optimized. By extracting structured data first, analyzing themes second, and synthesizing third, we produce summaries that are more complete, accurate, and useful than single-pass approaches.