Agent rules for 3D design projects

Working rules for designing a 3D model with an AI agent (Claude Code or similar) driving CAD software: who decides, how the model is built, how changes are recorded and documented.

Language	File
English	AGENT_RULES_3D.md
Français	REGLES_AGENT_3D.md

SolidCraft + GitHub + Claude Code (French, shorter, tool-specific): REGLES_SOLIDCRAFT_GITHUB_CLAUDE.md.

What they are for

An agent that drives CAD software works fast, but leaves a model that is hard to follow: changes nobody approved, unnamed features, values hidden in expressions, decisions that only exist in a past conversation. These rules keep the human in charge and the project readable:

• the agent proposes, the user approves, and nothing changes without approval;
• the model stays editable by hand: dimensions in sketches, named parameters, explicit names, built from a skeleton;
• every change is described in a design file, which is the readable record of a binary 3D model;
• it works with Git, a PLM or a plain file manager.

How to adopt them

1. Copy the rules file of your language into the project root, unchanged.
2. Create AGENTS.md with one line: "Before any action, read AGENT_RULES_3D.md in full (working rules), then DESIGN.md (project state, decisions, tools and pitfalls)." Claude Code reads CLAUDE.md: if your agent does not read AGENTS.md, add a CLAUDE.md that points to it.
3. Create DESIGN.md with the nine chapters listed in §3 of the rules (Goal, Input data, Model parameters, Model architecture, Decision log, Project tools and pitfalls, Open points, Next steps, History).
4. Fill in with the agent, before modelling: the goal, the input data, the software version table and the version control tool (chapter 6).

Limits

• Tested on a single project (a parametric staircase, SolidCraft + Claude Code), over a few days. Some rules have not yet been through a real change cycle.
• Not validated by a study. The few existing studies on agent rule files disagree on how much they help; they suggest keeping the file short, putting the most important rules first, and writing it yourself rather than letting an AI generate it.
• Written mostly with an AI, then directed and reviewed by a human.
• The agent's structural calculations are orders of magnitude, never a validated calculation.

Feedback and pull requests are welcome.

License

MIT