AI Usage Policy

Use of general-purpose language models (LLM)

As shared in the Brickheart development blog, general-purpose models (Claude, ChatGPT, Deepseek, Copilot) have been used to conduct research, help the author develop ideas further, generate sections of code, proofread writing and translate passages of text.

These models have deliberately been constrained:

• memory is disabled
• artefact generation is turned off (so all text appears in the chat)
• the models have no direct access to the project files
• Conversations are opted out of for training

As a result the models only know what the author chooses to share in that chat. For any new feature in the project, a chat session is first used to define the problem in detail, refine how it can be solved and plan the execution. The model then outputs a session summary containing only the agreed final plan.

After human review and editing of the summary, the feature is built in a separate chat from this and supporting files. When a bug is encountered the author first seeks to solve it themselves. If the issue is non-trivial, requires a moderate code revision or is not understood, the author requests edits only. For complex bugs or those needing a major code revision, the model is instructed to provide the patches or complete rewritten files.

For LLM output to be included in this project, the author actively takes the text and incorporates it. This is slower than giving the models direct write access to the project files but allows understanding of how the project's components fit together.

The code completion functionality in VS Code is used to help maintain appropriate syntax when making small changes to the code.

Use of LLMs for translation

One of the striking elements of LEGO® design is its lack of written language in the model instructions. This opens up the build experience across age and language barriers. The project has been constructed so that the main pages of the project website and accompanying app can be readily machine translated into a range of languages.

The aim of this feature is to broaden the experience beyond those fluent in English. The translations should be of sufficient quality to allow a user to understand, navigate and use the app in their preferred language.

Use of LLMs for writing assistance

This page was written by a human; I hope you can hear my voice. LLMs were used beyond basic spelling and grammar to highlight phrases which will translate poorly and use of overly complex language.

The following text is based on Clay's AI Writing Policy which takes a balanced approach to this technology. Its aim is to ensure people communicate well and make best use of each others time. I found it to be a helpful articulation of my own feelings.

Good writing is a tool authors use to clearly communicate ideas to someone else.

The author must own every sentence and idea in a piece of writing. It should represent their thoughts and place emphasis where they intended. When asked "What did you mean by ...?", a response of "The AI wrote that." shows a lack of ownership.

The writing process is a method of thinking that helps you engage with a topic. It requires time and serves as a 'proof of thought'. The artefact is only part of the goal, you understanding the topic by writing about it. Outsourcing writing to AI risks arriving at the artefact without any idea of if it contains anything worth sharing.

Respect other people's time by spending more time writing it than they will reading it. AI writing has a tendency to contain many sentences that don't say much at all.

AI doesn't fully understand what you as an author were trying to say to your readers. Asking it to rewrite, rephrase or summarise will result in meaning being lost. Imperfect prose that conveys the author's voice is preferable to 'polish'.