# Showing our work: a controlled experiment in context engineering

DevFeed: [Showing our work: a controlled experiment in context engineering](<https://devfeed.tech/articles/showing-our-work-a-controlled-experiment-in-context-engineering-101825.md>)

Original publisher: [Read original article](<https://www.getdbt.com/blog/context-engineering-experiment-dbt-mcp>)

Author: Christine Berger

Published: 2026-10-26T16:39:00Z

Content type: article

Language: en

Sources: [dbt Labs](<https://devfeed.tech/sources/dbt-labs.md>)

Topics: [AI-Ready Data](<https://devfeed.tech/topics/ai-ready-data.md>), [context](<https://devfeed.tech/topics/context.md>), [Retrieval Augmented Generation (RAG)](<https://devfeed.tech/topics/retrieval-augmented-generation-rag.md>), [workflow automation](<https://devfeed.tech/topics/workflow-automation.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-skills](<https://devfeed.tech/tags/ai-skills.md>), [analytics](<https://devfeed.tech/tags/analytics.md>), [analytics-engineering](<https://devfeed.tech/tags/analytics-engineering.md>), [api](<https://devfeed.tech/tags/api.md>), [context-engineering](<https://devfeed.tech/tags/context-engineering.md>), [cost](<https://devfeed.tech/tags/cost.md>), [experiment](<https://devfeed.tech/tags/experiment.md>), [insights](<https://devfeed.tech/tags/insights.md>), [sources](<https://devfeed.tech/tags/sources.md>), [time](<https://devfeed.tech/tags/time.md>)

## AI overview

In an 80-run test of one sales PDF skill, dbt MCP retrieval of full transcripts cost about 14% less and ran faster than Gong MCP while producing similar quality; warehouse filtering by account avoided Gong's retrieval limitations. Gong-generated briefs were cheapest and fastest, but across 20 runs missed contextual dates, numbers, titles, decisions, and next steps, and named only one of five discussed tools. A lightly engineered Snowflake summary retrieved through dbt MCP cut cost 48.5% and run time 43.2% versus the control, retaining most dates, numbers, and next steps and matching full transcripts on tooling. It captured titles less often (9/20 versus 19/20 for full transcripts) and missed decisions in all 20 runs. The authors conclude that modeled warehouse context can be reused across skills, while their extraction rules need improvement.

## Source excerpt

We ran one AI skill 80 times against four context sources. Engineered summaries cut cost 48.5% and run time 43.2%.