Rebuilding Historical Architecture with Claude Code: The Open-Source Monument Skill Delivers Interactive 3D Construction Timelines

Give an AI the name of a historical building, and let it automatically dig through historical records, estimate the real dimensions, and generate an interactive 3D model broken down by construction phase. What might sound like a request only an architectural historian would make has now been turned into an extension skill for Claude Code.

Developer Gleaming0427 has published an open-source project called Monument on Dev.to. Its usage is quite straightforward:

/monument Leaning Tower of Pisa, Italy

After entering this command, Claude will complete a full research and modeling workflow in sequence.

What Does Monument Actually Do?

Unlike "just asking AI to draw the Leaning Tower of Pisa," Monument's execution path is broken down into traceable steps:

1. Retrieve historical records about the building, sorting out its construction era and background;
2. Find its real footprint and architectural dimensions;
3. Identify the different construction phases;
4. Generate the building's form using procedural 3D geometry;
5. Generate an interactive timeline to observe the building's "stacking" process block by block;
6. Compare the result against reference photos and refine the model's accuracy accordingly.

This workflow deliberately places "research" before "generation." The developer emphasizes that each construction phase has a corresponding era and source citation, geometric shapes are prioritized based on verifiable real measurement data, and for details with insufficient information, the system explicitly marks them as "uncertain" rather than presenting them as established fact.

This means Monument's positioning isn't a visual toy, but rather an architectural history visualization tool with traceability.

Procedural 3D Geometry refers to generating three-dimensional forms automatically through code, parameters, and algorithmic rules, rather than drawing them face by face in manual modeling software. Traditional 3D modeling often relies on artists manually pushing and pulling vertices and sculpting details—suitable for artistic creation but difficult to scale, and inconvenient for directly converting textual information like "dimensions" and "era" into models. Procedural geometry is the opposite: developers first define parameterizable geometric rules such as footprint contours, story heights, tapering curves, and colonnade bays, then the program stacks tower shafts, domes, arches, and other structures layer by layer according to those rules. Its advantages are traceable dimensions, splittable phases, and reproducible results—as long as the input historical measurement data differs, the generated form changes accordingly. Monument chose this path precisely so that "which era each layer corresponds to, and which source it's based on" leaves a trace on the model, rather than generating a beautiful shell that can't explain its origins.

Technical Implementation: Local Operation, Three.js-Driven

Monument runs entirely locally, with 3D rendering powered by Three.js. The developer has currently produced three interactive reconstruction cases: the Leaning Tower of Pisa, the Eiffel Tower, and the Roman Colosseum.

The project is open-sourced under the MIT license, with installation via:

npx skills add Gleaming0427/monument

Beyond Claude Code, this skill also supports Codex, OpenCode, and Antigravity.

Project address: the Gleaming0427/monument repository on GitHub; interactive demos can be viewed at gleaming0427.github.io/monument.

Claude Code Skill is an extension mechanism Anthropic provides for Claude Code, allowing users to package a fixed set of instructions, tool-calling workflows, and resource files, triggered by a single slash command. It is essentially not a new capability trained into the model, but a structured "operation manual": after reading the skill definition, the model will sequentially call tools for retrieval, computation, file reading/writing, browser rendering, and more according to preset steps, compressing tasks that would otherwise require multiple rounds of conversation into a single command. Precisely because of this, the same skill can be ported to environments like Codex, OpenCode, and Antigravity that support similar extension protocols. Monument is one such skill package—it solidifies the chain of "check historical records—verify dimensions—split into phases—generate geometry—build timeline—compare photos," so users don't have to manually repeat the orchestration each time.

An Honest Limitation: Slow, and Requires Iteration

The developer proactively mentions the shortcomings of this approach: a fully detailed historical reconstruction can take 30 to 90 minutes, and the first generated result often requires manual fine-tuning.

This time cost isn't surprising. Procedural geometry generation itself relies on repeated validation, and adding historical record retrieval, phase identification, and comparison-based refinement against reference photos, each link lengthens the total duration of a single task. For scenarios requiring quick output, Monument isn't suitable; it's better suited for research or educational use where users are willing to wait for traceability.

From a broader perspective, Monument represents a direction of Claude Code skills that is taking shape—no longer limited to code generation and productivity tools, but stringing AI's retrieval, reasoning, and creative capabilities into a workflow with citations, to tackle more specialized and more "niche" knowledge-based tasks. The developer himself also posed a question in the post: which building should be reconstructed next?

Traceability in knowledge-based AI workflows refers to the ability for every step of a conclusion to point back to specific sources and evidence. Monument implements this through three concrete practices: construction phases labeled with era and source, geometry prioritizing verifiable measurement values, and details with insufficient information explicitly marked as "uncertain." Together these three form an "auditable" output—users not only see a tower, but can also know which year each layer was added, which proportion comes from which survey, and which details are actually speculation. This contrasts with the tendency of purely generative models to "confidently fabricate": the latter's failures often go undetected, while traceable workflows expose uncertainty, letting researchers, teachers, or students judge credibility for themselves. The 30-to-90-minute duration is largely the price paid for this audit chain.

One-line Summary

Monument weaves architectural history research and 3D reconstruction into a traceable Claude Code workflow.

Paragraph Summary

Monument is an open-source Claude Code extension skill by developer Gleaming0427. Enter a building's name, and it automatically retrieves historical records, verifies real dimensions, identifies construction phases, then uses procedural geometry to generate an interactive 3D model with a block-by-block "stacking" timeline. Unlike directly asking AI to draw, it places research before generation, labels each phase with era and source, and explicitly marks under-documented details as "uncertain"—making it more like a traceable architectural history visualization tool. The project runs entirely locally via Three.js, has reconstructed the Leaning Tower of Pisa, the Eiffel Tower, and the Roman Colosseum, is open-sourced under the MIT license, and is compatible with Codex, OpenCode, and Antigravity. The trade-off: a full reconstruction takes 30 to 90 minutes, and first results usually need manual tuning—better for research or teaching than quick output.