Why Tavily's Dynamic Filtering post answers its LinkedIn ads on substance, but loses the number-led hook in the H1

Tavily's LinkedIn cluster sells outcomes: 12x cheaper than Anthropic's Programmatic Tool Calling, 77.9% F1 versus 50.8% on the same model, 3.5x fewer tokens, and a single skill file that works in any coding agent. Every one of those claims is restated, sourced, and visualized inside the destination blog post.

The friction is at the top of the page. The H1 reads 'Dynamic Filtering: Let the Model Program Its Own Search Filters,' which names the mechanism instead of the result the ads were sold on. A visitor who clicked the cost-and-accuracy hook has to scroll past the abstract before the matching numbers appear. The audit lands at B+ because the substance is unusually strong but the hero copy does not echo the dominant ad promise.

The post is a 4-minute engineering write-up that introduces the principle of dynamic filtering, explains how Anthropic's Programmatic Tool Calling implements it, and then shows how Tavily rebuilt the same pattern as an open skill on the Tavily CLI that works in any harness with Bash and Python.

The benchmark section delivers the numbers the ads sell. On the full 900-question DeepSearchQA set the skill approach scores 72.7% F1 against Anthropic's reported 59.4% on the same Claude Sonnet 4.6 model. On a 50-question head-to-head run the skill scores 77.9% F1 versus 50.8% for PTC, averages 413K tokens per query at $0.37 against PTC's 1.46M tokens at $4.65, and lands at roughly 3.5x fewer tokens and 12x cheaper.

The article then explains why the pattern works - variable space stays in the execution environment, only print() output crosses into context - and closes with a 'Replicating it with the Tavily CLI' section and a 'Get started' call to action, which is what the ads' Learn more CTA is funding.