Strategy-Aware Summarization

Context

Summry's gating pipeline now classifies videos before summarization. Metadata creates an early prior, transcript quality checks source reliability, and final eligibility decides whether the video should continue into chunking and summarization.

Gate 5 uses the processing strategy produced by those gates. The goal is not to change the public summary format yet, but to make the prompts pay attention to the shape of the video.

Problem

Before this gate, every supported video used the same generic chunk-note and summary-reduction prompts. That meant a tutorial, lecture, interview, commentary piece, and review all received the same summarization focus.

The summary contract was stable, but the prompt had no way to emphasize the details that make each video type useful for watch decisions.

Implementation

This gate adds strategy-aware prompt instructions inside the summarization service.

The summarizer now accepts:

• processing strategy
• eligibility status

The worker passes both values from the persisted video gate state. The prompt then adjusts its focus:

• tutorials prioritize prerequisites, steps, tools, caveats, and failure modes
• explainers prioritize concepts, mechanisms, examples, and limitations
• lectures prioritize definitions, dependencies, progression, and evidence
• interviews prioritize claims, disagreements, topic shifts, examples, and takeaways
• analytical videos prioritize thesis, arguments, assumptions, evidence, and implications
• reviews prioritize criteria, tradeoffs, evidence, pros and cons, and who should watch or skip

Caution videos also receive a conservative uncertainty instruction so the model avoids inferring details that are not supported by the transcript.

Tradeoffs

The main benefit is better alignment between video type and summarization focus without changing the response contract.

The tradeoff is that prompt routing is still heuristic. It depends on earlier gates choosing a useful processing strategy. The summary format remains unchanged, which keeps the UI stable but limits how category-specific the output can become.

Lessons Learned

The simplest first step was to route instruction focus, not redesign the summary shape. That keeps this gate reviewable and reduces the risk of breaking existing Markdown rendering.

Prompt specialization can now improve incrementally. Later work can decide whether some strategies deserve different section templates, but the current implementation keeps the contract steady.

Next Steps

The next useful step is evaluation. Run the same videos through the baseline and strategy-aware prompts, then score watch-decision value, specificity, coverage, and accuracy before changing the output structure.