Architectural Prompting: does it add up? The skill that checks the consistency between project documents

The problem: each paper is correct, but together they don’t add up

A project is made up of documents written at different times and often by different people: the technical report, the metric calculation, the economic framework, the timetable, the specifications, the safety plan. In the meantime the project changes. A process is added, the customer requests a variation, a price is updated, and the change enters the document that is open at that moment.

Thus the amount of the work in the economic framework differs by a few thousand euros from that of the report, the specifications describe a 14 centimeter coat and the calculation accounts for a 12 centimeter one, the timetable lasts 180 days and the specifications allow 150. Taken one by one, the documents are in order. The error lies in the comparison, and it is the type of error that in a tender becomes a request for clarification and in the construction site a challenge to the company.

Finding it requires reading all the documents while keeping dozens of numbers in mind. An AI assistant does it well, if you give it a method and prevent it from doing its own thing.

What the skill does

The skill receives the documents of the same intervention and extracts from each one the data that also appears in the others: amounts, quantities, durations, dates, processes, names, references to the tables. It compares them in pairs, redoes the sums and multiplications that it can read and returns a table of findings, each with the two conflicting values ​​and the page or entry in which they are found.

The reliefs are of three types. Discrepancy indicates two different values ​​for the same data or a calculation that doesn’t add up. The absence indicates a process described in the specifications and absent in the calculation, or vice versa, or a table referred to which is not there. The third level, “to check”, collects the comparisons that the assistant was unable to make, for example because a table was illegible.

I built the skill around a prohibition: the assistant doesn’t decide which of the two values ​​is right and doesn’t correct anything. It tells you where to look. If he chose him, you would find yourself with a coherent package and perhaps aligned on the wrong number.

It does not even check whether the design choices are correct or compliant with the regulations. It’s a consistency check, the one that in the studio is entrusted to the colleague with the most finicky eye.

How to install

1. Open your chatbot (Claude, ChatGPT, Gemini).
2. Drag the SKILL.md file found below into the chat and ask to install it as a skill.
3. Wait for confirmation from the assistant.
4. To use it, upload the documents and write, for example: “Check the consistency between these documents before delivery”. The assistant will understand on his own that he needs to use the skill.

If your assistant doesn’t yet manage the skills, paste the text of the file into the instructions of a custom GPT, a Project or a Gem: it works the same way, only you have to recall it yourself.

Instructions and suggestions for use

1. Upload all submissions to the same conversation. Control thrives on comparisons and a document left out is an uncontrolled document.
2. Add the processed list, if you have it. The skill uses this to tell you what is missing and to verify table recalls.
3. Prefer native files to scans. On a poorly scanned calculation the assistant misreads the figures, and the skill requires him to declare it, but a half-done check remains half-done.
4. On very long calculations, randomly check the calculations redone by the assistant. The models still make mistakes in the sums with many rows, and for this reason the table always shows the written value next to the recalculated one.
5. Repeat the check after each correction. Fixing an amount in one document is the most common way to create a new discrepancy in another.
6. The documents contain names, addresses and cadastral data. Before uploading them, check the settings of the tool you use, in particular if the contents are used for training.

A final note

This skill is deliberately generic. The next step is to adapt it to your study: open the file with a text editor and add to the list of key data those that in experience escape you most often, whether they are the resistance classes of the concrete or the acronyms of the fixtures. After two or three projects you will have your checklist, written once and applied the same way every time.

Disclaimer

This skill is a review support tool and does not replace the verification of the project by the professional. The assistant may not detect a discrepancy or misread a piece of data: the absence of findings is not equivalent to validation. The final check of the documents remains the responsibility of the technician who signs them. Happy prompting!