8. AI Assisted Development#
AI tools like GitHub Copilot[1] can greatly enhance productivity, but effective use requires a few fundamental practices. The following practices are essential for reliable and efficient AI-assisted coding.
8.1. Core Rules for Effective AI-Assisted Coding#
Preserve Essential Context
Record important prompts, code snippets, and decisions outside the AI session (e.g., in comments or project notes), since session context is not persistent across restarts or between editor windows.
Engineer Explicit Instructions
Formulate clear, specific instructions[2] to guide the AI toward relevant and accurate suggestions.
Independently Verify AI Output
Critically review all AI-generated code for correctness, security, and adherence to project standards before accepting or merging it.
Independently here means verifying the output yourself, without relying on (or being unconsciously influenced by) the AI’s own assurances or using the AI to check its own work.
Remember: Copilot can help you write tests, but cannot run them. Always run and check yourself.
Redirect Context When Necessary
AI tools lack the ability to recognize when a course change is necessary. It is the responsibility of the developer to identify when the current approach is no longer effective or optimal and to redirect the AI accordingly.
Note
While each instance of redirecting context may seem small, the cumulative effect of such deliberate shifts can lead to what Thomas Kuhn described as a “Paradigm Shift”—a fundamental change in approach or perspective. In this sense, redirecting context is not just a practical action but a critical part of driving innovation and progress in development workflows.
Each of these axioms is fundamental and cannot be derived from the others. By following them, you ensure effective and reliable use of AI in your coding workflow.
8.2. How to Save Your AI Session Context#
Since chat and prompt history in VS Code (and similar environments) can be lost when you close a window, change workspaces, update extensions, or restart your machine, it is important to proactively save essential context.
Recommended Practice:
After a productive session with GitHub Copilot, ask the AI to summarize the key context, prompts, and decisions from your session. Save this summary in a file named COPILOT_CONTEXT.md (or another descriptive name) in your project directory.
This way, you can easily recover your working context after a restart or when switching between environments, ensuring continuity and minimizing the risk of losing important information.
Example:
At the end of your session, prompt:
“Summarize the context of this session to ensure the continuity of this chat for the next session.”Save the AI’s summary in
COPILOT_CONTEXT.md.
By following this practice, you make your workflow more robust and resilient to interruptions or environment changes. This workflow helps ensure you never lose important context, even if your environment changes.
8.3. How to Recover Your AI Session Context#
To resume work after a restart or interruption, use the COPILOT_CONTEXT.md file (or another descriptive name) in your project directory to restore your working context.
Recommended Practice:
Open the
COPILOT_CONTEXT.mdfile.Review the summarized key context, instructions, and decisions from your previous session.
Prompt the AI:
“Recover the context from
COPILOT_CONTEXT.mdto ensure the continuity of this chat.”
By following this process, you can quickly restore your workflow with minimal loss of continuity, even after interruptions or environment changes.
8.4. Managing Formatting Inconveniences in the VS Code Chat Pane#
8.4.1. Recognizing the Inconvenience#
When working with GitHub Copilot in the VS Code chat pane, you may encounter formatting issues—such as broken or split code blocks, misrendered markdown, or incomplete responses. These problems often arise when sharing code, markdown, or quoted text, making it difficult to copy, reuse, or understand Copilot’s suggestions.
8.4.2. Why These Issues Occur#
GitHub Copilot does not have direct access to how its responses are rendered in your VS Code editor. It cannot see if a code block is broken or if markdown is misinterpreted. Instead, Copilot relies on your instructions and its training data. As a result, formatting issues can persist until you explicitly point them out or set clear preferences.
8.4.3. Using RESPONSE_PREFERENCES.md to Avoid Issues#
To minimize these inconveniences, you can create and maintain a RESPONSE_PREFERENCES.md file in your project. This file should clearly state your formatting preferences and any known issues to avoid. For example, you might specify:
Always provide code or markdown in a single code block using four backticks and the appropriate language tag.
Do not split suggestions into multiple blocks unless explicitly requested.
Place explanations after the code block.
Include file paths as comments when relevant.
By reminding Copilot of these preferences at the start of each session (or when issues arise), you help ensure more consistent and usable responses in VS Code.
8.4.4. Resetting Preferences to Defaults#
GitHub Copilot does not automatically detect or use the contents of RESPONSE_PREFERENCES.md unless you explicitly instruct it to do so in your prompt. If you want to return to the default Copilot response style, you should clearly state your intention in your prompt. For example:
“Please ignore the RESPONSE_PREFERENCES file and use your default response style from now on.”
“Reset to default response preferences.”
Removing or renaming the RESPONSE_PREFERENCES.md file alone has no effect unless you also update your instructions to Copilot. Always be explicit in your prompts when you want to change or reset response preferences.
By proactively managing your response preferences, you can avoid many common formatting issues and make your Copilot-assisted workflow in VS Code smoother and more productive.
Note
The use of a context reminder file (such as RESPONSE_PREFERENCES.md) is a practical approach that can be generalized to address other types of recurring inconveniences in AI-assisted workflows. Given the current state of tools, it is necessary to explicitly remind the AI of your preferences or known issues, as so-called “feedback” provided in conversation is not persistently respected or remembered by the AI. Always use clear, explicit instructions and context files to ensure your preferences are followed.
8.5. How to Change Context During a Session#
If your goals, requirements, or focus change during an ongoing session, or if you realize that the current approach is no longer effective, it is your responsibility to redirect the AI. Many AI assistants, including GitHub Copilot, rely on the current file and prompts for context. To ensure relevant assistance, explicitly guide the AI toward the new direction.
This section focuses on general context changes, such as switching files or adding new information to the current session. For example:
“I am now switching to
utils.py. Please disregard the previous context.”
Or, if you want to keep some previous context but add new information:
“In addition to the previous context, now consider that I am working on a new feature in
main.py.”
Tip
If a file or task is no longer relevant to your current focus, consider closing it to avoid distractions and maintain a clear working context. In addition, pay special attention to the current file, because it is treated in a special way as mentioned above.
While this section addresses general context changes, Rule 4: Redirect Context When Necessary focuses on specific scenarios where deliberate redirection is required to address limitations in the AI’s approach. These scenarios often arise unexpectedly and require the developer’s judgment to intervene. For more details, see the section “Necessary Context Redirection under Rule 4.”
Being clear about context changes and the current file helps the AI provide more accurate and relevant suggestions as your work evolves.
For further tips or more detailed guidance, feel free to ask GitHub Copilot directly as you work. This section is intentionally kept simple to encourage direct engagement with AI assistants for further guidance.
8.6. Necessary Context Redirection under Rule 4#
While general context changes are common during development, Rule 4 (Redirect Context When Necessary) addresses specific scenarios where the need for redirection often arises unexpectedly. AI tools, like GitHub Copilot, may persist in iterating on an ineffective approach, attempting to refine a solution even after repeated failures.
In such cases, it is the developer’s responsibility to recognize when the current approach is no longer viable and to intervene by redirecting the AI toward a more effective path. This process is not always anticipated in advance but requires awareness and judgment to identify when a shift in focus or strategy is necessary.
8.6.1. Common Scenarios for Context Redirection#
Switching to Library Utilization:
When to Redirect: If you start with a basic implementation (e.g., using plain HTML/CSS) but later realize that a library (e.g., D3.js) would be more efficient or powerful.
How to Redirect:
Ask the AI for recommendations on libraries that align with your goals. The AI often has a broader knowledge of available tools and can suggest candidates you might not be familiar with.
Once you’ve identified a suitable library, inform the AI about your choice and its purpose.
Provide relevant documentation or examples to guide the AI in generating solutions using the library.
Example Prompts:
To explore library options:
“What JavaScript libraries are available for creating interactive data visualizations? Please suggest some options and briefly describe their strengths.”
To redirect after choosing a library:
“I want to switch to using D3.js for this task. Please disregard the previous plain HTML/CSS approach and focus on generating D3.js-based solutions.”
Introducing Unit Testing:
When to Redirect: During development, if the AI struggles to handle complex code changes or repeatedly fails to make progress, it may be necessary to simplify the problem and verify the approach through unit testing or a minimal example. This redirection is required because the work is ongoing and cannot proceed effectively without intervention.
How to Redirect:
Clearly state the need to simplify the problem or focus on testing specific scenarios.
Suggest creating a simplified version of the code to verify the AI’s understanding and approach.
Once the simplified code works as expected, ask the AI to apply the solution to the original, more complex code.
Example Prompts:
To focus on testing:
“The current implementation is incomplete. Let’s write unit tests for the key scenarios to verify the logic before proceeding further.”
To simplify a complex problem:
“The current code is too complex for this change. Let’s create a simplified version of the function to verify the logic. Once it works, we’ll apply it to the original code.”
To apply the simplified solution:
“Now that the simplified version works, please use it as a reference to update the original, more complex code.”
Dividing and Ordering Focus:
When to Redirect: If a task becomes too complex, it may be necessary to break it into smaller, more manageable parts and address them sequentially.
How to Redirect:
Define the specific subtask you want to focus on first.
Clearly communicate the order of priorities to the AI.
Example Prompt:
“Let’s focus on implementing the data parsing logic first. Once that’s complete, we’ll move on to the visualization component.”
8.7. Refining Practices and Terminology#
This chapter itself demonstrates how to refine practices and terminology collaboratively with an AI assistant. Here’s how the process unfolded about the first section above.[3]
Legend:
🧑💻 Actions I took during the process.
🤖 Contributions made by the AI assistant.
Start with an Initial Proposal:
🧑💻 I prompted the AI to propose general practices for effective AI-assisted coding wihout any specifics.
🤖 The AI generated a list of five practices, including ideas like “managing session context” and “reviewing AI-generated code.”
Simplify and Structure:
🧑💻 I asked the AI to simplify the proposed practices into a smaller, more fundamental set of principles. To guide this process, I introduced the term “axiomatic” to emphasize the need for foundational and independent rules.
🤖 The AI suggested combining the practices into three “axioms,” which were:
Preserve Essential Context
Engineer Explicit Instructions
Independently Verify AI Output
💡 Note: At this stage, Rule 4 (Redirect Context When Necessary) was not yet included. It was later identified as a missing principle during the refinement process, as described in Step 4.
💡 Tip: Thoughtful word choice, like starting with “axiomatic,” can significantly enhance clarity and focus during refinement. Choose terms that convey your intent precisely and minimize the need for additional explanation.
Iterate and Refine:
🧑💻 I reviewed the “axioms” and realized that while the term “axiomatic” effectively conveyed the foundational nature of the rules, it might feel too formal or mathematical for some readers. I consulted the AI to refine the terminology further.
🤖 The AI suggested replacing “axioms” with “Core Rules” (including other options) to make the language more familiar and accessible. It also helped rephrase specific rules for clarity and usability.
Example: The AI suggested clarifying “Preserve Context” as “Save key prompts and decisions in project notes to ensure continuity.”
💡 Tip: Refinement is an iterative process. Be open to revisiting terminology and structure to ensure they resonate with your audience while maintaining precision.
Identify Missing Principles:
🧑💻 While reviewing the initial three axioms, I realized that they did not address a critical limitation of AI tools: their inability to recognize when a course change is necessary.
🤖 The AI assisted in brainstorming how to address this gap, leading to the idea of a new rule focused on redirecting context during significant course changes.
Add Rule 4:
🧑💻 I added Rule 4, titled “Redirect Context When Necessary,” to address this gap. The rule emphasizes the developer’s responsibility to identify when the current approach is no longer effective and to guide the AI accordingly.
🤖 The AI contributed actionable examples, such as updating prompts and providing additional context, to make the rule practical and easy to apply.
Improve Readability:
🧑💻 To make Rule 4 more engaging and relatable, I proposed a note referencing Thomas Kuhn’s concept of a “Paradigm Shift” to help readers understand it more deeply.
🤖 The AI suggested phrasing the note to connect the practical action of redirecting context with the broader philosophical idea of driving innovation and progress.
💡 Tip: Adding contextual notes or analogies can help readers connect abstract principles to real-world applications, making the content more memorable and impactful.
By following this process, the Core Rules evolved into a more comprehensive and balanced framework, addressing both the strengths and limitations of AI tools like GitHub Copilot.
8.8. Final Note[4]#
AI-assisted development is a rapidly evolving field, and the practices outlined in this chapter are designed to help you make the most of current tools like GitHub Copilot. However, as AI tools and development environments continue to improve, it’s important to remain adaptable while asserting your preferences and goals without compromising easily.
The collaborative process described in this chapter—generating ideas, simplifying them into actionable principles, and iterating for clarity—can be applied not only to AI-assisted coding but also to other areas of your work. By combining thoughtful human input with the capabilities of AI, you can create workflows that are both efficient and resilient.
Finally, remember that while AI can assist with many aspects of development, your judgment and expertise remain essential. Use AI as a tool to enhance your productivity, but always verify its output and ensure it aligns with your goals and standards.
By staying engaged, adaptable, and assertive in your preferences, you can continue to refine your practices and make the most of AI-assisted development as the field evolves.
Note
The practices and limitations described in this chapter are based on the status of AI tools (such as GPT-4 or GPT-4.1) and Visual Studio Code as of 1.100. Capabilities and behaviors of AI assistants and editors may change in future versions.