Architectural Prompting: from the AI ​​that draws to the AI ​​that controls. How to create a reviewer of project documents

A project dedicated to revision

Both ChatGPT and Claude allow you to create gods Projects: work environments in which to collect reference documents, conversations and standing instructions. Availability and features depend on the plan used.

Unlike a prompt inserted into a casual conversation, a Project’s instructions define reusable behavior. In our case, we can configure the AI ​​to act as an assistant for document review, maintaining the same analysis criteria and return of results. To get started simply:

1. Create a new Project, in ChatGPT or Claude, named for example Reviewer of technical documents.
2. Enter the standing instructions below, adapting them to your needs.
3. Add reference documents, such as product nomenclatures, control checklists and study conventions.
4. Start a conversation by uploading the versions of the documents to compare and specifying the required verification.

It is advisable to dedicate different conversations to different activities: comparing revisions, checking the graphics and checking consistency between documents.

Prompt of the Week: The AI ​​Auditor

The following text constitutes an initial configuration, to be copied into the Project instructions. It is not designed as a universal procedure: each firm should integrate it with criteria, documents and checks specific to its activity. Instructions to start from and copy into the Project:

ROLE AND OBJECTIVE
Act as a technical assistant for the review of construction project documents. Your task is to compare documents and reviews, identify differences, possible omissions and inconsistencies, preparing a register of checks to submit to the designer.
You are not a substitute for professional technical inspection and should not automatically treat a difference as an error.

ANALYSIS METHODS
1. Document comparison
When you receive two revisions of the same document:

• Locate added, deleted and modified content.
• Highlights variations in quantities, dimensions, materials, performances and regulatory references.
• Reports the position of the changes, indicating page and paragraph when identifiable.
• Distinguish confirmed changes from possible inconsistencies.

2. Graphic comparison
When you receive two versions of a plan, section, elevation, or detail:

• Identify the documents and verify that they are comparable in scale, orientation and content.
• Analyze differences in geometry, openings, distribution, dimensions and legible annotations.
• If image processing tools are available, try performing an alignment and graphical overlay of the boards, producing an image of the differences.
• Distinguish design variations from those due to layout, resolution, or export quality.
• Do not declare the inspection complete and do not obtain precise measurements from images without reliable dimensional references.

3. Cross-check
When you receive different documents of the same project, look for possible contradictions between graphic information, descriptions, quantities and technical specifications. Report necessary checks without making automatic judgments of regulatory compliance.

RETURN OF RESULTS
Produce an observation log containing:

• Code and revision of the paper
• Page, paragraph or area affected
• Description of the difference or anomaly
• Type of report
• Proposed priority (high, medium, low)
• Verification requested from the designer

Always distinguish verified elements, uncertain ones and non-analysable ones. At the end it specifies which checks have been performed and which require further manual verification.

CUSTOMIZATION
Adapt the procedure to the instructions and reference documents provided by the practice. If criteria necessary to carry out an audit are missing, report them before proceeding.

From documents to drawings: what can we expect?

Comparing two reports or specifications is a relatively simple starting point: the system can spot textual changes, compare numerical values, and organize the differences in a table.

More attention is required with graphic designs. Multimodal models can recognize obvious changes, such as a different internal distribution or the addition of an opening, but do not guarantee to detect every geometric or elevation change.

For a first experiment it is advisable to use two tables exported in high resolution PNG, keeping the scale, framing and dimensions identical. A graphical overlay procedure, performed with image processing tools, can help locate variations that the AI’s visual readout alone might overlook. The assistant can then describe and classify the differences identified, while the dimensional and geometric verification remains entrusted to the designer and the appropriate technical tools.

ChatGPT, Claude or other tools? The test is part of the method

Which platform works best? There is no one-size-fits-all answer and performance may change with model updates. The suggestion is to set up the same reviewer in multiple platforms and compare the results using a pair of documents with already known changes. You can start with two reviews of a technical report and continue with two versions of the same plan.

The evaluation should consider not only how many changes are detected, but also the misreporting, the accuracy of the references and the ability to recognize one’s limitations.

Before uploading professional documents, it is also necessary to check the data processing conditions, the confidentiality settings and the possible need to anonymize the documents.

The goal is not to replace the technician’s control, nor to fully replicate the functions of specialized software. Rather, it is to build an assisted review procedure that is repeatable and progressively adaptable to the practice’s activities. The quality of the result will depend not only on the model used, but also on the control criteria that the professional will be able to define.

References and insights