Why can't I just upload my drawings to Claude or ChatGPT?

A full drawing set is usually too large for a chat upload, and even when it loads, the model pushes every sheet into its context window at once. Accuracy drops as the context fills, a problem called context rot. Converting the set to text first avoids both the upload limit and the accuracy loss.

What is a drawing.md file?

A drawing.md file is a plain-text description of one drawing sheet: its contents, material specifications, dimensions and cross-references. AI creates one per sheet by querying an embedded copy of the set. Later tasks read the text file instead of the PDF, which uses around 30 times less context.

Can AI measure dimensions or quantities off construction drawings?

No. AI reads and extracts what is written on a drawing, including text, tags and schedules, but it does not understand spatial relationships and it cannot reliably scale or measure. Counts need a human verify pass, and any measured dimension or quantity stays with the estimator.

Do I need coding skills to set up a drawing database?

No. Google Antigravity is a free desktop agent, and the Airtable MCP connection is a small block of configuration code that AI will write for you. You paste in your Airtable personal access token, open the agent in your drawings folder and run the extraction prompt.

What happens when the drawings get revised?

Re-run the indexing on the revised sheets and update your drawing register. Mark superseded sheets rather than deleting them, the same discipline you would apply to any document control loop. A stale index is worse than none, because AI will answer confidently from out-of-date information.

Which AI model is best at reading construction drawings?

Google's models have historically been stronger at visual analysis, which is why Gemini and NotebookLM feature in both workarounds. Claude's recent Opus models have improved noticeably on visual information. The preprocessing matters more than the model choice, because every model performs better querying text than interpreting images.

AI for Reading Construction Drawings: A How-To

Getting AI to read a full set of construction drawings is one of the first walls contractors hit. You try to upload the set, the upload fails or the answers come back vague, and the whole idea starts to look overhyped. This guide is for builders and estimators who want drawing answers they can check and trust. It comes from my own testing on a live warehouse drawing set, in the video below.

Key takeaways

A raw PDF drawing set is close to the worst input you can give AI: big files, dense symbols, and constant cross-references between sheets.
The fix is preprocessing. Convert the set into text once, then have AI query the text instead of re-reading the PDFs.
Method 1: extract the vector data from every sheet into a searchable Airtable database using Google Antigravity, then connect Claude to it.
Method 2: upload the set to NotebookLM and have Claude Code write a drawing.md text file for every sheet.
The line does not move: AI reads, extracts and indexes drawing information. It does not understand spatial relationships and it does not measure.

Can AI read construction drawings?

AI can read construction drawings reliably only after the set has been converted to text. Upload a raw PDF set into a chat and it fails or answers vaguely. Convert each sheet into structured text first, in a database or a drawing.md file, and AI answers drawing questions quickly and accurately, because it is querying text rather than interpreting images.

These models are trained text in, text out. Processing a chunk of text is cheap and accurate for them. Processing a dense A1 sheet full of symbols, line types and references to other drawings is not. There is a useful detail hiding inside that problem though: most PDF drawings carry a vector data layer, the embedded text along with its positions and orientations. That layer extracts programmatically at near-perfect accuracy. Vision is for context. The vector data is where the accuracy lives.

So the job is not to make AI better at looking at drawings. The job is to hand it the text layer in a form it can search.

Why do drawing sets break AI chat uploads?

A typical set is simply too big for a chat. My warehouse test set was 25 drawings at 10.7 megabytes, and it failed to upload to both the Claude chat and a Claude Project. Even when a set does load, the model pushes every sheet into its context window at once, and accuracy drops.

That accuracy loss has a name: context rot. The more information you give a model in one go, the less likely it is to answer correctly. On paper the context window is big enough for a drawing set. In practice, jamming 25 sheets into it produces slow, vague and sometimes wrong answers.

The third route people try is Cowork, pointing Claude at the folder holding the drawings. It runs, but Cowork reads the whole PDF set and uses around 30 times the tokens of the equivalent text. On a Pro subscription you can burn most of your usage working through a single set. And you are still asking a model that is not the strongest visual analyser to do the visual processing. Opus 4.7 was a real step up on visual information, and I had been leaning on Gemini for drawings before that, but preprocessing beats both options.

Method 1: turn the drawing set into a searchable database

The first workaround indexes the whole set into a database that AI queries on demand. A desktop agent splits the merged PDF into individual sheets, extracts the vector data from each one and writes it to Airtable. Claude then connects to the base over MCP and pulls only the sheet data a question needs.

Here is the setup I used:

Put the drawing set in one folder. Mine was a merged 25-sheet PDF.
Install Google Antigravity. It is a free desktop agent, think Claude Cowork for the Google ecosystem, downloaded from antigravity.google. Open it inside the drawings folder.
Connect Airtable over MCP. MCP is just a way of connecting AI tools together. In Antigravity’s MCP settings you paste a small block of configuration code with your Airtable personal access token (in Airtable: your account, then Builder Hub, then personal access tokens). AI will write the configuration code for you if you ask.
Run the extraction prompt. Mine tells the agent to split the merged set into individual drawings, convert each one to an image, extract the vector data from every sheet, and write small Python scripts to work through those steps one at a time. Small incremental steps are where these agents behave.
Write the results to Airtable. Each sheet becomes structured records: what is on it, the text and tags, the references to other drawings.
Connect Claude to the same base and ask questions. “Using my Airtable connection, give me a summary of the warehouse drawings” now searches a text database instead of re-reading 25 PDFs.

The reason this works is that you have stopped asking the model to find a needle in the whole set. As I put it in the video:

“It’s like giving someone a set of 30 or 100 drawings and saying, ‘What is the height of the retaining wall?’ versus giving someone the retaining wall drawing and telling them this is the height of the retaining wall.”

One note on the database choice: Google Sheets works too, Excel does not. When Claude reads an Excel file it pulls the whole workbook into context, which recreates the original problem.

Method 2: NotebookLM and drawing.md files

The second workaround is quicker to set up and leans on Google’s stronger multimodal reading. Upload the set to NotebookLM, which embeds documents well, then have Claude Code query the notebook and write a drawing.md text file for every sheet. Every later task reads those small text files instead of interpreting the original PDFs.

The steps:

Upload the drawing set to NotebookLM. It is Google’s research tool, built for students, but it handles drawing sets surprisingly well.
Connect NotebookLM to Claude Code. The integration is a little finicky the first time, but the setup is simple once you follow a guide, and there are good walkthroughs on YouTube.
Open Claude Code in the drawings folder and link the notebook.
Run the drawing.md prompt. The goal I give it: for every drawing, create a drawing.md file capturing the contents, material specifications, dimensions and cross-references, written so that someone doing a task like preparing an estimate or a program only needs the .md file, not the sheet.
Point your later work at the text files. Ask Cowork or Claude a question using the drawing.md files and it reads a small text document instead of processing a PDF.

The efficiency difference is the headline number from my testing: on average, a drawing.md file uses around 30 times less context than reading the equivalent PDF drawing. Faster answers, less usage, more accurate output, because the model works from a small, relevant chunk of text.

This is a preprocessing step, and it pays off across everything downstream. An indexed set is the starting point for Claude for construction estimating, and the same text layer keeps answering questions during delivery, the territory covered in AI for construction contract administration: checking a variation against the current revision, or pulling drawing references into an RFI.

Where the line sits: what AI still cannot do with drawings

AI reads, extracts and indexes what is written on a drawing: title blocks, text, tags, schedules and specifications. It does not understand spatial relationships between elements, and it does not measure. Any scaled dimension or quantity stays with a human. Treat extraction output as a draft to verify, not a measurement to trust.

Drawing task	AI fit	Notes
Drawing register from title blocks	Good	Pure extraction; verify status codes on a sample
Indexing sheets into text (drawing.md)	Good	Run once per set, re-run per revision
Text, tag and schedule extraction	Good	Specs, finish schedules, reinforcement schedules
Counting items off plans	With care	Approximate counts only; needs a human verify pass
Scaling or measuring dimensions	No	The human measures
Coordination and clash judgment	Limited	AI can surface overlaps from the index; who moves is your call

The reason for the hard line on measurement is that vision models do not read a drawing the way we do. A dashed line means something completely different to a bold line, and a small change in a symbol changes the meaning entirely. Spatial reasoning is exactly where these models are weakest, and counts come back approximate, not accurate. That is why the takeoff question gets its own answer in can AI do a construction takeoff: AI sets up the count, estimates the count or checks the count. It does not do the counting.

None of this makes the indexing work less useful. It just means the index answers “where is it and what does it say”, and you answer “how big is it and how much”.

Common mistakes to watch for

Uploading the raw merged set to a chat and trusting whatever comes back. If the model even accepts the file, you are in context-rot territory from the first question.
Using vision for questions the vector data already answers. The embedded text layer is near-perfect when extracted programmatically. Keep visual analysis for context, not for facts.
Letting the index go stale through revisions. Re-run the indexing when sheets are revised, keep the drawing register current, and mark superseded sheets rather than deleting them. A confident answer from a superseded drawing is worse than no answer.
Storing the index in Excel. Claude pulls the whole workbook into context. Airtable, Google Sheets or plain markdown files hold up better.
Asking AI to measure because the extraction went well. Extraction and measurement are different jobs. The first is text retrieval; the second is spatial judgment.
Skipping your own read of the set. The index makes AI faster at answering questions. It does not replace knowing the job. You still read the general notes, orient on the grids and levels, and understand what you are building.

Both methods are working experiments rather than finished systems, and I am still refining the prompts and the data structure. The packaged versions, the drawings analyser and project indexer skills, are in the ContractorOS community. Watch the full video above to see both of them run end to end on the warehouse set, including the failed uploads and the exact prompts I used.