Can AI Write an Architect's Snagging List From Site Photos?

Yes, as a draft; no, as an issued document. AI can turn site photos and walkround notes into defect items grouped by room, each with a suggested severity and responsible party. The architect must still check every item against the photo, the specification and the contract before issue, because the practice signs the list, not the software.

The reason is what a model can and cannot see. It describes what is in the frame, and both OpenAI and Anthropic note that image models give approximate counts and can misread rotated, low-quality or very small images. It cannot know your finishes schedule, the tolerance the specification allows, what was agreed on site last month or what you did not photograph. A shadow becomes a crack, an unfinished edge becomes a defect and a design decision becomes the contractor's fault.

Follow me on Instagram@sagnikteaches

A snagging list is also a professional statement. It tells the contractor what must be put right before practical completion (or substantial completion, depending on the contract), and a wrong line can start a dispute or a claim. That makes the checking routine the real subject, with the AI doing the typing, sorting and first wording.

Connect on LinkedInSagnik Bhattacharya

The example running through this page is invented but typical. A small practice has just walked a single-storey rear extension and kitchen remodel on an occupied family home, and has 46 photos, a page of spoken notes and five areas to cover.

Subscribe on YouTube@sagnikteaches

Where the AI's work ends and the architect's begins

The middle column is typing and sorting, and the right-hand column is judgement. Keep the two apart and the AI saves real time without being able to embarrass you. Let them blur, and a confident sentence from a model starts to read like an inspection.

The tutorial on an architect's site visit report from notes and photos uses the same capture habits for ordinary progress visits. A snagging list differs in one way that matters: every line is an instruction to a contractor, so the cost of a wrong line is higher.

Shoot the walkround so the photos can be read later

The quality of the draft is decided on site. A photo that shows the defect, its location and its scale needs little interpretation, and a photo that shows only a close-up of a white surface needs a lot.

• Three frames for anything that matters. One wide shot that shows the room and where the defect sits, one medium shot of the element, and one close-up with a tape or coin in view for scale.
• Raking light for paint and plaster. Hold a torch or phone light at a shallow angle across the surface. Dust, roller marks and ridges that vanish in flat light show up at once.
• A number and a spoken note for every photo. Start each note with the room and the element, then the observation: "K-03, kitchen, worktop, mitre joint open at the front, no sealant."
• Upright, uncropped, in focus. OpenAI's guide says images may be resized and that rotated images can be misread, and Anthropic's says very small or low-quality ones produce mistakes. A crooked, blurred shot is a guess for you and for the model.
• Check every frame for people and paperwork. On an occupied home, a photo can include a child, a letter on a table or a screen. Delete or crop those before anything is uploaded.

Upload limits shape the batches. On claude.ai you can attach 20 images to a message at up to 10 MB each. ChatGPT accepts images up to 20 MB each, with up to 80 file uploads every three hours, and its Free plan allows only three file uploads a day, which makes a 46-photo list impractical in one sitting on that plan. Batching by room avoids both problems. In the worked job the 46 photos fall into five batches: kitchen-diner 18, utility 6, cloakroom 6, hall and stairs 8, exterior 8.

The photo-note sheet that does the matching

Before drafting, build one table with a row per photo: its number, the room, the element and the tidied note. Paste the table with the photos for one room, and the AI can join evidence to observation instead of guessing which photo is which.

The note column is the most valuable one. The AI should treat it as your observation and the photo as evidence, and when the two disagree it should say so instead of choosing. That single rule prevents most of the quiet errors described below.

The drafting prompt, and what comes back

One prompt per room keeps the table short enough to check. This one bans causes and blame, and makes disagreement visible:

You are helping an architect draft a snagging list for a domestic extension. I've attached the photos for one room and a table of photo numbers, rooms, elements and my spoken notes.

For each photo, write one defect item: a short, factual description of what is visible plus what my note says. Do not name causes, do not say who is to blame, and do not describe anything that is not visible or in my note.
If the photo and the note disagree, or the photo is too dark, small or blurred to judge, write CHECK and say why.

Suggest a severity using my scale (1 = safety, weathertightness or use; 2 = function or finish, fix before handover; 3 = minor or cosmetic) and a responsible party from this list: main contractor, named subcontractor, supplier or manufacturer, client's own supplier, design team query.

Output a table: item number, room, element, description, photo number, severity, responsible party, action needed.
Finish with a list of every photo that produced no item and every note that has no photo.

Five rows of a typical reply, combined across rooms for illustration:

Three of those rows need the architect.

• Item 2 is the rule working. The note said "check against lighting plan" and the photo could not settle it, so the AI wrote CHECK instead of inventing a verdict. The architect looks at the lighting layout and either deletes the item or confirms it.
• Item 5 is the danger. The AI blamed the contractor and gave it severity 1, because a step at a door sounds serious. The drawing specifies a 40 mm upstand at that door, so this is a design question about access, not a build defect. Issued as drafted, the practice would have told the contractor to rip out work built to the architect's own detail.
• Item 4 holds a guess. The AI wrote "decorator" because paint was involved. Nobody gave it the contractor's subcontractor list, so the trade is invented. Use "main contractor" unless you know who painted that hall.

The fix for item 5 goes into the prompt for next time: "If an item relates to something shown on the drawings as designed, mark it as a design team query unless my note says the work differs from the drawing." AI tends to default to the contractor, so the instruction has to say otherwise.

Severity and responsibility: set the scale before the visit

A severity scale is a practice decision, and it must exist before the walkround, not be improvised by a model afterwards. The three-level scale in the prompt is one workable version. Level 1 covers safety, weathertightness and anything that stops a room being used, and it holds up practical completion. Level 2 covers function or finish that is clearly wrong and must be fixed before handover. Level 3 covers minor or cosmetic items that can be closed within an agreed period after handover. Add a query flag for anything that might be a design matter rather than a defect.

Responsibility is about who must act, not who is to blame. The five categories in the prompt (main contractor, named subcontractor, supplier or manufacturer, the client's own supplier and design team query) are enough for a small job. The client's own supplier matters on a kitchen, where the client may have bought the appliances or the worktop separately. The supplier or manufacturer category covers a faulty product rather than faulty fitting. A model cannot tell these apart from a photo, and only you know what the contract says about each.

The tutorial on a snagging report template covers the report layout, severity order and photo captions in more detail. Paste your own scale into the Project instructions once, so every room starts from the same definitions.

The line-by-line check before issue

This is the step that makes the list yours. Work down the table with the photo open beside each item, and do not skip the ones that look fine.

1. Read the item cold. Cover the AI's description, look at the photo, and decide what you see. Then compare.
2. Compare with your note. Models drift towards stronger words: "slightly" becomes "significantly" and "check" becomes "defective".
3. Test it against the documents. Is it a defect against the specification, a tolerance the specification allows, or design intent? The finishes schedule and the drawing detail decide, not the photo.
4. Confirm the location. Right room, right element, right side. Left and right are easy for a model to swap.
5. Confirm severity and responsibility. Re-grade anything that does not match your scale, and change any party you cannot support from the contract.
6. Strip opinion. Remove causes, blame and remedies you did not decide.
7. Merge duplicates. The same defect photographed twice becomes two items unless you merge it.
8. Look for what is missing. Compare the list with the finishes schedule and your memory of each room. Sealant, ironmongery, touch-up paint, cleaning and testing are the usual gaps, because they are rarely photographed.
9. Refresh anything that may have changed. If days passed between the visit and the check, confirm that disputed items are still true.
10. Record who checked. Initial the final version, and keep the draft and the marked-up copy together.

Wording needs the most care, because a spoken note and a polished sentence can mean different things. A spoken note might say "skirting's a mess round the stairs". An AI draft turns that into "Skirting around the stair is poorly finished and must be replaced", which prescribes a remedy nobody chose. The issued line should read "Skirting at the foot of the stair: gap at the mitre and exposed filler; make good and repaint." It states what is wrong and leaves the method to the contractor.

Existing conditions need their own check. If the home was photographed before work began, compare the item with that record, as the tutorial on a schedule of condition drafted from photos describes. A scuff on a retained staircase may be older than the contract.

Where models misread site photos, and what to do about each

The same few mistakes recur, and each has a cheap defence at the time of capture or in the note.

• Glare and reflections read as wet patches. A glossy tile or window can look like a leak. Take one frame with the light source behind you, and say "reflection" in the note if you saw one.
• Shadows read as cracks and gaps. A dark line beside a skirting may be a shadow, a shrinkage gap or a pencil mark. If the description mentions a crack, a gap or a line, make the AI write CHECK unless your note says what it is.
• Size from a photo is a guess. "About 2 mm" belongs in your note, taken with a rule or tape in frame. Tell the AI to use only the measurements you gave it.
• Counts are approximate. "Several chipped tiles" can become four or eleven. Count on site and put the number in the note.
• Colour and gloss cannot be judged from a phone. A shade difference between two coats of paint may vanish or appear depending on the light, so record what you saw in the room and leave the judgement to your own eyes.
• Structure, fire protection and electrics are not for the model. A crack that might be structural, a missing fire stop or a suspect circuit should be written as "refer to the structural engineer" or "refer to the electrical inspector", and never classified or explained by the AI.

None of these makes photo-based drafting unsafe. They explain why the note column in the sheet carries so much weight, and why the architect's own observation always outranks the picture.

A second pass: asking what each room should contain

A draft list from photos can only describe what was photographed, so the biggest risk is silence. A second prompt turns the finishes and fittings schedule into questions that jog your memory:

Here is the finishes and fittings schedule for the cloakroom (text) and the draft snagging table for the same room. List every schedule item that has no matching defect item and no photo, as a question for me: "Did you check [item]?" Do not assume any item is fine or defective.

For a cloakroom, a typical reply asks whether you checked the extractor fan's operation, the door lock and handle, the sealant between basin and wall, the toilet seat fixings, the ceiling paint finish and the window cill. In the worked job, three of those six questions turned into items and the rest were confirmed fine. The prompt does not find defects. It finds the places where you have not yet said anything, which is where an incomplete list hides.

A small practice's pass: 46 photos in five batches

Here is how the worked job played out. The five batches produced 41 draft items, with three marked CHECK and five photos that produced no item. The architect's review changed the list as follows.

The arithmetic is 41 draft items, less 5 deleted, plus 6 added, which makes 42 issued. The AI work took about ten minutes across the five batches. The rest of the effort was human: 25 minutes building the photo-note sheet, 75 minutes on the check and 20 minutes on layout and issue, so 130 minutes from the end of the walkround to the issued list. Suppose the practice's last comparable list took five hours to write up by hand. The saving is nearly three hours, and almost none of it comes from the review, which is why the review should never be the part that gets squeezed.

Six of the 42 issued items existed only because the architect checked the list against memory and the finishes schedule. Without that step the AI list would have looked complete and been wrong about the room.

Issuing the list and closing items out

The issued document is a PDF with a numbered table: item, location, description, photo thumbnail, severity, responsible party, action and target date. The AI can draft the cover note and lay out the table, and you adjust the wording to match the contract's own terms for practical completion, outstanding items and the period allowed to rectify them. Date it, give it a revision number, and send the contractor a copy that matches the one you keep.

The same tool helps again when items are closed out. At the re-inspection, photograph each item again under the same numbering with an R1 prefix, then ask the AI to compare the new photos with the issued list:

Here is the issued snagging list and the photos from the re-inspection, numbered R1 plus the original photo number. For each item give a status: COMPLETE (the new photo clearly shows the fix), NOT COMPLETE (the defect is still visible), or CANNOT TELL (the photo does not show enough). Quote the item number and describe what you can see in the new photo in one sentence. Do not judge quality against the specification.

Treat CANNOT TELL as a request to go back and look, and confirm every severity 1 and 2 item in person before you mark it closed. The status table is a working aid for the architect and a record for the file, but it is not a certificate of anything. Keep each revision of the list, because disputes later in the defects period often turn on what was listed and when.

Client photos and privacy: what not to upload

Photos of an occupied home hold more than defects. They can show family members, post, screens, valuables and security equipment, and the client agreed to a building project, not to their home being fed into a chatbot. Delete or crop anything personal before upload, name files with job codes rather than addresses, and leave client names out of the prompt.

Then check the plan you are using. The team plans, ChatGPT Business and Claude Team, keep uploads out of model training by default; a personal plan has its own model-training switch in privacy settings, and setting it is your job. Read it before the first upload, and write the answer into the practice's policy. The tutorial on an architecture firm AI policy for drawings and client data gives a checklist, and the one on what to tell your professional indemnity insurer covers the insurance side of the same question.

None of this makes the snagging list less useful. It changes who does the typing. The architect still walks the building, still decides what is a defect and still signs the list, while the AI takes the sorting and first wording off the desk so the checking gets the time it needs.

Questions architects ask before using AI on snagging

Should the contractor be told that AI helped draft the list?

The list goes out under the practice's name, so you are accountable for every line however it was drafted. Whether to mention the tool is a question for your appointment terms and your insurer. Ask them, write the answer into the practice's AI policy, and keep a record of who checked each item before issue.

Can voice notes replace typing during the walkround?

Yes, if every note starts with the photo number, the room and the element, for example 'K-03, kitchen, worktop joint'. Dictate on your phone, tidy the text into the photo-note sheet, and paste that sheet into the AI. The sheet matters more than the audio, so check each line against the photo before drafting.

How many photos can I upload at once?

On claude.ai you can attach 20 images to a message at up to 10 MB each. ChatGPT takes images up to 20 MB each and allows up to 80 file uploads every three hours, a cap OpenAI may lower at busy times, while Free accounts get three uploads a day. Either way, batch by room, which also keeps each draft table short enough to check.

What if the AI misses a defect that is in the photo?

It will sometimes, because models give approximate counts and skim detail. Your walkround notes stay the primary record, and every photo should map to an item or be listed as producing none. Ask the AI to name any photo without an item, then look at each of those yourself.

Further reads

• AI in the Construction Field: Photos, Voice Notes and Punch Lists — See how photos, voice notes and punch lists work for people building the job.
• Writing Client Updates and Snag Lists With AI for Builders — Read the builder's side of snag lists and client updates.
• Architect Meeting Minutes: Record, Summarise and Issue Same Day With AI — Issue site meeting minutes the same day with the same checking habit.
• Copyright and Client Data Risks of AI in Architecture Practices — Understand copyright and client-data risks before uploading project material.
• Claude vs ChatGPT vs Gemini for Architects: Which to Pay For — Choose which assistant to pay for before building a routine on it.
• AI Policy for Architects: A Ready-to-Use Checklist and Clauses — Write the practice policy that decides what staff may upload.
• AI for a Painting Business: Snag Lists and Final Walkthrough Sign-Off — See the contractor's view of a snag list and final sign-off.
• AI Tools for Architects 2026: A Toolkit From Brief to Handover — Which AI tool to use at concept, design, tender, site and handover, what each costs on its own pricing page, and three stacks for a five-person practice.
• AI for Architecture: Keep a Drawing Issue Register and Transmittal Log — One long-form register, AI that turns issue emails into checkable rows, formulas for current revisions and a ten-point check before anything is sent.
• ChatGPT for Building Design: Briefs, Options and Code Checks — Brief intake, layout options and a code-check log for a small design practice, with the line ChatGPT must not cross.
• Claude (Formerly Cowork) for Architects: Five Admin Tasks to Try First — Five folder-based admin jobs for a small practice, from an actions roll-up to a fee check, with prompts, sample output and what to verify.
• Glide Mobile App Builder: A Snag List Photo App for a Building Firm — Which Glide to pick, how to build a sheet-backed snag list with camera photos and status filters, how to show clients only their plot, and the plan cost.
• Practical Completion Certificate: What Each Clause Means, Then an AI Draft — Each clause of a practical completion certificate in plain English, a filled-in kitchen extension example, and an AI draft you check against the contract.
• Structural Engineer Inspection Report: Photos and Notes to a Draft With AI — A photo-to-report workflow for structural engineers: numbered photos, a register, AI descriptions of what is visible, and checks before the engineer signs.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI's image input guide and its help page on file uploads; Anthropic's vision documentation and its help article on uploading files. The practice, photos and figures in the worked example are illustrative.