Skip to content

1. Getting Started

To get started working with AI ways of working, it's helpful to have a conceptual understanding of how to use this playbook to generate quality, consistent code.

Workflow overview

Below is a diagram that provides a high-level overview of the steps defined in this playbook. For the most part, this development workflow follows recognised best practices already established throughout government, with the additional aid of AI tools and techniques.

---
config:
  layout: dagre
---
flowchart TD
    A["Product Idea"] -- Prompt with input --> B["Functional Requirement"]
    B -- Prompt with input --> C["Technical Specification"]
    C -- Prompt with input --> D["Tests"] & E["Application Code"]
    D -- Test passes --> E
    E -- Refactor code as necessary --> D
    E -- Prompt with input --> G["Create / update documentation"]
    G -- Prompt with MCP --> H["Code review on Pull Request"]
    H --> E
    H -- Approve pull request --> I["Feature Branch Merged"]

Image: Simplified development workflow diagram

Workflow stages

  • Product / Service idea - This represents the problems to be solved for your users using existing service design and user research techniques.
  • Prompt to generate functiional requirement - Using advanced models, such as the latest "thinking" models, the clearly defined ideas can be used to generate requirements documentation (features, user stories, data models, etc.) that the AI IDEs (Integrated Development Environment - Cursor) can later use to generate code.
  • Prompt to generate technical specification - The requirements document is then used to create a detailed technical specification that defines the architecture, implementation approach, and technical details needed for development.
  • Prompt to generate tests and application code - Both tests and application code are generated from the technical specification in the AI IDE, ensuring the business logic defined in the requirements are tested independently from the code generation.
  • Test-driven development cycle - Tests must pass before code is considered complete, with refactoring prompted as necessary to ensure code quality and test compliance.
  • Prompt to create / update documentation - Documentation can be kept up to date by prompting the AI IDE to update documentation based on the changes that have been made.
  • Create Pull Request for developer code review using MCP - A Pull Request (PR) in git is then generated following traditional development practices, with Model Context Protocol (MCP) tools assisting in the review process. Each line of code is reviewed for quality and brevity, ensuring that the code to be deployed is production-ready.
  • Merge feature branch - Once the PR is approved and merged into the main branch, the feature branch is merged and automated pipeline processes are used to deploy the code, as per your normal deployment processes.

Generating Consistent Code

To generate consistent code, multiple elements must come together as depicted in this diagram:

Image: Venn diagram depicting the intersection of elements for consistent AI code generation

Element Purpose Where created
Clear Requirements Defines (with detail and clarity) the functional and pseudo-technical requirements for the product/service idea you want to implement. Using an advanced "thinking" chat model like Claude or ChatGPT.
- Ref: Product Requirements Workflow
Good Prompts The clear, detailed ask for a given task. It can refer to the IDE rules and the requirements, however, it is not necessary to repeat what's in the rules or requirements in the prompt. Prompts are initially created manually but then can be further refined using a chat model like Claude or ChatGPT.
- Prompt engineering and meta prompting is explained further in the playbook: Prompting Guidance
IDE Rules for AI IDE rules define consistent and repeatable standards, patterns and conventions across the codebase. Rules can be applied for each task-specific prompt. Rule file formats are usually defined by the AI IDE tool. To aid in generating the rules themselves, a chat model like Claude or ChatGPT can be used.
- Ref: Language-specific playbook rules files
Capable Code Generation Model Using the most capable LLM model for the task you are running is important for good quality results. Not all tasks require advanced models, so selection of the most cost-effective model that can achieve the desired outcome is ideal. The AI IDE tools typically allow the user to select which model is used when prompting the LLM.
- Currently, the latest Claude Sonnet models are recommended starting models for quality code generation.

Next steps

To get started, ensure that your project is set up for AI development and follow the steps in the Playbook Workflow. IMPORTANT - this includes ensuring that privacy settings are enabled, per the project is set up section.

We also advise reading and understanding the detailed Prompting Guidance before you start.