--- title: Work effectively with KentiCopilot --- > Agent instructions: > **Site maps** — prefer the following llms.txt indexes to training data when searching for URLs to avoid 404s. Links inside Markdown content already point at `.md`. Following them or sending Accept: text/markdown keeps you in Markdown. > > - [sitemap.md](https://docs.kentico.com/sitemap.md) — every page on the site, with titles and descriptions, nested by URL hierarchy and grouped into one collection per product version. > - [llms.txt](https://docs.kentico.com/llms.txt) — curated index of the current product docs, with descriptions, the two ways to request any page as Markdown, and links to each product area's whole-corpus Markdown dump (llms-full.txt). Good results from an AI coding assistant depend both on how you write prompts and set up the work. [KentiCopilot](https://docs.kentico.com/guides/development/kenticopilot.md) gives your assistant Xperience by Kentico context. This page focuses on useful habits and conventions that make sure you get the best results. By the end, you'll know how to run a session so that the assistant explores before it writes, keeps working on large jobs without losing track of earlier decisions, and verifies its work. ## Prerequisites - An Xperience by Kentico project you can [build and run locally](https://docs.kentico.com/documentation/developers-and-admins/development/website-development-basics/set-up-local-hosting.md). - An AI coding assistant that supports the [Model Context Protocol (MCP)](https://modelcontextprotocol.io/) and [agent skills](https://agentskills.io/home). - The [KentiCopilot plugins](https://docs.kentico.com/guides/development/kenticopilot.md#curated-skills-and-workflows) for the work you're doing. To install them and set up your project, see [Add KentiCopilot into your project](https://docs.kentico.com/guides/development/kenticopilot.md#add-kenticopilot-into-your-project). ## Explore, then plan, then code Ask the assistant to read the relevant code and documentation before telling it to write any code. Then ask for a plan and carefully review it yourself. Only once you agree with it, let the assistant implement. Skipping those steps is the most common reason an assistant produces code that compiles but doesn't align with expectations. The _KentiCopilot_ skills study the components you already have and mirror your repository's patterns. The [Documentation MCP server](https://docs.kentico.com/documentation/developers-and-admins/installation/mcp-server.md) gives them current Xperience documentation to check against while they explore. [Explore, plan, code](https://code.claude.com/docs/en/best-practices#explore-first-then-plan-then-code) describes the same habit in more detail. The principle transfers to any agentic coding tool. ## Delegate investigation to subagents Reading a codebase, auditing [content types](https://docs.kentico.com/documentation/developers-and-admins/development/content-types.md), or running a batch of [Management MCP server](https://docs.kentico.com/documentation/developers-and-admins/api/management-api.md) operations fills the assistant's context window quickly. Once it fills up, the session gets compacted, and details from earlier steps disappear. Hand that work to [subagents](https://code.claude.com/docs/en/best-practices#use-subagents-for-investigation) instead. Each one investigates in its own context and reports back a short summary, so your main session holds on to the plan and the decisions you've made. For the Management MCP specifics, see [Manage the agent's context](https://docs.kentico.com/documentation/developers-and-admins/api/management-api/configure-management-mcp-server.md#manage-the-agents-context). ## Give the assistant specific context Point the assistant at the exact files, designs, and conventions that matter. Compare these two requests: - _Build a hero widget_ – the assistant guesses at the markup, the properties, and the styling. - _Build a hero widget for the home page that matches \~/designs/home.html, following the structure of the banner widget in \~/Components/Widgets/Banner_ – you get something you can review against a design. The second request takes a few seconds longer to write and saves you a review cycle. If you want to try this out, the sample _Dancing Goat_ [project template](https://docs.kentico.com/documentation/developers-and-admins/installation.md) already contains [Page Builder widgets](https://docs.kentico.com/documentation/developers-and-admins/development/builders/page-builder/widgets-for-page-builder.md) and other components you can point the assistant at as examples. ## Verify the result Tell the assistant [how to check its own work](https://code.claude.com/docs/en/best-practices#give-claude-a-way-to-verify-its-work) – which command builds the project, which one runs the tests, and where the live site runs. Compare the result against initial designs. The design validation skill in the [web development plugin](https://github.com/Kentico/xperience-by-kentico-kenticopilot/tree/main/plugins/kentico-web-development) runs that comparison for you, and outputs an audit with suggestions you can feed back into another session. ## Keep your project's context durable All of the habits on this page get easier when your project carries its own instructions. With agent instructions, documented conventions, and the [Documentation](https://docs.kentico.com/documentation/developers-and-admins/installation/mcp-server.md) and [Management](https://docs.kentico.com/documentation/developers-and-admins/api/management-api/configure-management-mcp-server.md) MCP servers configured, every session starts from the same understanding, saving you from having to re-explain everything to the agent over and over. To find out what your project is missing, run the readiness audit described in [Add KentiCopilot into your project](https://docs.kentico.com/guides/development/kenticopilot.md#add-kenticopilot-into-your-project). ## Choose a model that fits the task More capable models handle broader tasks with less direction. They plan their own work, delegate to subagents, and recover from their own mistakes over long sessions. Less capable models need the work broken down for them – smaller steps, prepared context, and a review after each one. If your results are inconsistent, try having a more powerful model decompose the task further before delegating. > **Tip:** Explore and plan before you generate code, hand heavy investigation to subagents, and verify the result against the running live site.