Claude skills for architects are reusable instructions for recurring work: reviewing a drawing register, reconciling schedule exports, or preparing a model change brief from project decisions. We have created a downloadable starter pack for those three tasks, with fictional input records and expected findings so your team has something concrete to evaluate.
The pack contains plain-text workflow instructions and JSON examples. It does not install a Revit connection, execute code, or change a model. Use it to make document review more consistent, then test the results against your own project rules.
What are Claude skills for architecture and BIM?
An Agent Skill packages instructions in a folder containing SKILL.md, with name and description metadata followed by the procedure. Supporting references can sit alongside that file. The format gives an assistant reusable task context; it does not grant application access or permissions. Agent Skills specification.
For a BIM team, the distinction between a skill and a connection is useful. A connection determines which data and tools an assistant can access. A skill describes how to approach the work, what ambiguities to preserve, and what output to produce. Anthropic’s guide describes skills as a way to package repeatable expertise and workflows, including workflows that use MCP tools. Anthropic’s skills guide.
A good drawing-review skill therefore does more than say “check this PDF.” It tells the assistant to retain source references, distinguish unreadable pages from missing sheets, and avoid turning a packaging issue into a model-edit instruction. The value lies in those decisions and their evidence, not in the length of the prompt.
Download the BIM workflow skills and sample cases
Download the complete BIM workflow skills pack. The ZIP contains a README and three separate skill folders, each with a SKILL.md procedure and supporting reference files. The examples are fictional JSON records, not customer documents or native Revit projects.
You can also download the individual packages: drawing-register review, schedule reconciliation, and BIM change brief. Each individual ZIP contains one skill and its references. Client import formats vary; inspect the folder structure and follow the instructions for your installed client.
We checked the package structure and manually reviewed the expected findings against the fictional inputs. We have not run a behavioral benchmark in Claude or a native Revit test for these skills. The sample expectations are a starting point for your evaluation, not a guarantee that an assistant will always follow every instruction.
Use a drawing-register review skill
The drawing-register example compares an expected sheet list with observations from a PDF package. It deliberately includes an unreadable page, a changed title, and two pages carrying the same sheet number with different issue labels.
A101 matches the expected register. A201 has a title discrepancy that needs review. A102 has not been identified in the readable pages, but one page is unreadable, so the evidence does not establish that A102 is absent. A501 appears twice with different issue labels; both observations must be retained until someone resolves the package history.
These distinctions change the next action. An unreadable page calls for a better source or a manual check. A duplicate PDF page may call for a package correction. Neither observation, on its own, authorizes deleting or renumbering a Revit sheet.
Run the example first and compare the assistant’s findings with the supplied expectations. Look for lost qualifications as well as wrong answers. “A102 missing” sounds clear, but it discards the uncertainty that makes the example useful. When adapting the skill to PDFs, ensure the extraction step supplies page coverage and legibility information; the sample JSON assumes those observations already exist.
Reconcile schedule exports and prepare a BIM change brief
The schedule example uses document identity together with an element ID. The same numeric ID appears in two different fictional models, so matching on ID alone would mix unrelated elements. One external record also has conflicting duplicate entries. The expected result preserves that conflict rather than choosing whichever row appears last.
Units and scope provide two more checks. A length of 1,200 millimeters matches 1.2 meters after an explicit conversion, while an empty field does not equal zero. The example export is partial, and one proposed value belongs to a type field. An assistant cannot establish every affected instance from that partial export alone.
Ask the skill for a reconciliation report and unresolved questions before considering an update. Include the actual export scope, itemization, field meaning, and units when adapting the procedure. A parameter label such as Mark should not be assumed unique, and a type-level change should not be described as affecting only the selected instance.
The change-brief example addresses a different failure: confusing recency with authority. An authorized project architect approves retitling sheet A201 while retaining its number. A later coordination participant suggests renumbering it. The expected brief preserves the approved retitle and treats the newer suggestion as unresolved. It does not mark either action as executed in the model.
Together, the skills support a useful sequence: inspect the source package, reconcile the relevant records, and prepare a brief containing only changes with established authority. For broader context, see our explanation of an AI BIM assistant.
Install, evaluate, and adapt skills for your team
Unpack the download and read each SKILL.md before importing it into a compatible client. Install the individual skills using that client’s current instructions. Connecting project data or native applications is a separate step; the pack itself supplies neither credentials nor an application integration.
For evaluation, provide one fictional input case and ask the assistant to follow the matching skill. Compare its report with the expected findings, including uncertainty, duplicate handling, unit conversion, and approval conditions. Preserve the actual response so a later prompt or model change can be evaluated against the same case.
Then adapt a copy to your firm’s naming rules, review roles, and export structure. Add representative examples with sensitive information removed and check that the instructions still work when a source is incomplete or contradictory. Avoid promoting a skill solely because it gives an attractive answer to its simplest example.
When the review produces an approved native task, ArchiLabs Desktop can support work in Revit and AutoCAD alongside the relevant documents. Keep the skill’s report attached to that task and verify the native result. Our guide to what AI can and cannot do in Revit helps separate document reasoning from model execution.
AI BIM Assistant streamlines design, clash detection, and documentation—accelerating BIM workflows, reducing errors, and improving project delivery.
Understand what Revit AI can do, where model access and source data limit results, and how to review geometry, parameters, and documentation.