Skip to main content

Recommendations for working with the agent

This page is useful when Explyt works without an explicit error, but the result is unpredictable or takes too many corrections. To diagnose a specific symptom, open the troubleshooting section.

Give a verifiable task

State the expected result and how to verify it: a specific test, a build, the absence of an error, or a change in behavior. For long work, describe the stop criteria explicitly.

Attach an entry point

Add the file, the selected code, the test, or the error message. Ask the agent to name the files and symbols its conclusions are based on.

Break down large tasks

Prepare a plan first, then run small steps. Start a new unrelated task in a new chat. You can send extra ideas for the current task through the message queue.

Lock in a repeatable approach

Keep project context in AGENTS.md, instructions for files in Rules, and repeatable scenarios in Skills. Do not duplicate the same instruction across all of these mechanisms.

Limit actions technically

Use Edit Scope, .readignore, .writeignore, and the Tools menu. A request in plain text helps, but it does not replace a technical limit.

Check the result

Review the changes in Agent Changes, run the tests or the build, and run Auto Review if needed. For a security audit, use vulnerability search.

Connect the sources you need

For an external knowledge base or service, use MCP. For internet search, use Web Search. To work with a user interface, connect a browser MCP server.