# Building with Coding Agents

> Use the kit PRD workflow to shape a feature, implement one story and review it before completion.

*Canonical: https://makerkit.dev/docs/void-cloudflare/development-guide/working-with-agents*

---

Open the repository in your coding agent with the `makerkit` MCP server connected. Ask for the outcome in plain words, for example: “Add organization-owned projects. Members can read them; admins can create them.” The agent reads `PRODUCT.md`, repository rules and the relevant docs before planning.

## The Feature Workflow

{% img src="/assets/images/docs/void-cloudflare/agent-workflow.svg" width="600" height="936" alt="Feature request, PRD shaping, architecture, implementation, review, then human acceptance. Agents stop at review." /%}

PRDs live in `.prds/` and are managed through MakerKit tools. Do not edit their JSON by hand. One functionality gets one PRD; a story is a small, verifiable behavior. Behavior-neutral corrections and refactors do not need a new PRD.

The names above are kit prompts/workflows, not shell commands. If your client does not expose MCP prompts, ask the agent to use MakerKit's `get_implementation_prompts` tool for the implementation instructions.

## Manage PRDs with Loop

Loop is MakerKit’s UI for PRDs. Use it alongside your coding agent to browse the plans in `.prds/`, inspect stories and track their progress without editing JSON by hand.

Start it from the kit repository:

```bash
pnpm run loop
```

The coding agent still implements and verifies each story. Use Loop to follow the work and review stories waiting for your acceptance; agents leave them in **review** until you mark them **completed**.

## Resume Work

Name the PRD and story you want to continue. Ask the agent to read `get_story` and its hand-off log before changing code. `get_project_status` reports progress; `get_next_story` finds work whose dependencies are completed. Review stories already waiting for you before starting their dependants.

A useful hand-off says what changed, what was tested, how to try it and what remains uncertain. Agents stop at **review**; a person marks **completed** through the PRD tools after checking the acceptance criteria.

## What Needs You

You decide product scope, try the delivered flow and accept stories. Setup may hand you a login, dashboard step or secret prompt; follow that one step and let the agent resume. Never paste secrets into chat. See [Coding Agent Permissions](../security/coding-agents) for exactly what tools can reach.

For verification commands and where tests belong, use [Testing and Debugging](./testing-and-debugging). A passing typecheck alone does not establish that a user can complete the flow.
