Databases, Custom Tools, & Workflow Automation

Test data, temporary UI controls, and documentation checks can make software easier to verify, but developers often skip the setup. Thorsten Ball, Ryan Carson, Graham McBain, and guest Colleen Brady show how they use agents to do that preparation and automate recurring development tasks.

Make the states you need to test easy to reach

• Seed a local database with deliberate scenarios. Ball asks Amp to create threads belonging to different users with different visibility settings so he can test a new starring feature. The agent initially guesses column names incorrectly, then inspects the schema and inserts the data. He also describes temporary UI controls that let a developer toggle billing-page states without constructing a separate test application 14:49–18:43(opens in a new tab).
• Use mock data while exploring a design. Brady iterates on a page using a TypeScript data file before moving the data into a database. This lets her settle the UI's data needs without repeatedly changing the database during design work 19:33–20:34(opens in a new tab).
• Start the next task from the diff. Ball gives a fresh agent conversation git diff or git show output. The diff supplies the changes and shows which files were edited together, helping the agent pick up related work without a long verbal reconstruction 23:13–24:16(opens in a new tab).

Automate a specific job, then expose it to the agent

• Choose a specialized extractor when it fits the data. Carson reports better results on financial PDFs with Google's Document AI than with his earlier general-model attempts. He describes defining the fields to extract, labeling examples, and calling the resulting processor through an API. His speed, cost, and accuracy assessments concern his workflow, not a comparison across all document types 25:07–30:03(opens in a new tab).
• Check documentation when committing. Brady says stale instructions repeatedly sent her agent back into the same mistakes. Her npm run commit workflow inspects changed files, helps prepare the commit message, and checks which documents need updating or archiving 32:59–35:41(opens in a new tab).
• Distinguish a user command from an agent tool. In the Amp version demonstrated, a custom slash command runs a script when the user invokes it and sends its output as a message. A toolbox exposes scripts that the agent can choose to call while working. Both reuse ordinary scripts; the difference is who decides when to run them 36:43–40:59(opens in a new tab).

Full Video: Watch on YouTube(opens in a new tab)

Build Crew: https://buildcrew.team(opens in a new tab)