Start with an outcome: which page, who it is for and what should change. Include the content and references the agent should use. Keep the first task small enough to review in one page.
Choose how to work
- MCP: use a local coding agent connected to Desktop.
- Chat with OpenAI: use the experimental ChatGPT connection in the local development build.
- Chat with OpenRouter: choose a model and use built-in contextual chat.
These routes have separate connections. Connecting an external MCP client does not configure Desktop’s chat provider.
Give the agent project context
In Desktop, use the project’s brief, Brand guide and connected Flow skills to
explain the site. For built-in chat, enable the relevant context before sending.
An external agent should read project, project_context and the relevant
flow and flow_skill results before authoring.
When working directly in the project, ask the agent to read its AGENTS.md
and the guidance bundled with the installed @blockbite/astro package:
npx --no-install blockbite context
npx --no-install blockbite context --listKeep existing project instructions. The CLI reads guidance; it does not rewrite your instructions or prove that an agent followed them.
Ask for a reviewable change
For example:
Update the home page introduction for first-time visitors. Use our existing Brand guide and supplied copy. Keep the section editable, then check the saved page in Astro at desktop and mobile sizes.
Review the proposed edits, saved files and native result. A successful save, a passing native check and a visual review each tell you something different. Share only the project context needed for the task; keep credentials out of prompts and project files.