Architect Site Visit Report: Notes and Photos to a Clear Report With AI

Number your notes and photos on site, give the AI only what you saw, and ask it to sort each item into observation, instruction and action columns, with a separate list of what you did not see. It drafts the wording and captions; you check every line against your notes before the report is issued.

The risk is not a clumsy report. It is a fluent one. AI writes in confident, finished sentences, and left alone it will turn "looks fine??" into "satisfactory" and add a membrane check you never made. A site visit report is a record that contractors, clients and, in a dispute, other people will read closely, so the whole method below is built around one habit: the assistant may rephrase what you observed and nothing else.

Follow me on Instagram@sagnikteaches

If you did not see it, the report cannot say you did

A site visit report records three things: what you observed, what you were told, and what you said or instructed. It never records an item as inspected because it was on the schedule, appeared in the last report, or would normally be fine. Reports are written from memory under time pressure, and AI makes gap-filling easier because it supplies plausible sentences for free. So the prompt has to forbid it, and your check has to look for it.

Connect on LinkedInSagnik Bhattacharya

That is why the structure matters more than the model. A report with separate columns for observation, instruction and action, and a separate list headed "Not observed", leaves the assistant nowhere to hide an assumption. The tutorial on snagging lists from site photos covers the defects walk-round; this one covers the periodic progress visit.

Subscribe on YouTube@sagnikteaches

Five status words the report is allowed to use

Give the assistant a fixed vocabulary and tell it to use no other. A restricted list of status words is the cheapest control you have.

"Reported" is the one people forget. Contractors tell you plenty on site, and some of it matters, but the report must show that it came from them. "Membrane laid before blockwork (reported by site manager, not seen)" reads very differently from "membrane in place".

Capture on site so the AI has something true to work from

The report is only as good as the notes, so the capture habits are worth ten minutes of thought.

• Start every set of notes with a header: visit number, date, arrival and departure times, weather, who was present, the purpose of the visit and the areas you intend to see.
• Number as you go. Write the photo number into the note when you take the picture ("P4 = first-floor joists, east side"). Reconstructing numbers at the desk is where captions go wrong.
• Take three kinds of photo for any item that matters: a wide shot that shows where you are, a medium shot of the element, and a close-up of the detail.
• Write certainty into the note. "Measured", "looked", "told" and "couldn't see" are the most useful words you can type. They are the words the assistant must preserve.
• Record the limits while you walk. "Didn't go up to the roof" takes three seconds on site and saves a doubtful sentence later.
• If you prefer to talk, say the numbers aloud. "Photo four, first-floor joists, east side, spacing looks wide, not measured" survives transcription far better than "this one here". The method is the one in turning on-site voice notes into job sheets.

When you upload photos, put the date and the location in your message. Anthropic's vision documentation says Claude does not parse or receive any metadata from images, so the time and GPS tags stored in the file never reach the assistant. The same documentation gives the limits on claude.ai: 20 images per message and 10 MB per image, so a 40-photo visit goes in two or three batches grouped by area. It also warns that Claude can make mistakes with low-quality, rotated or very small images, and that its counts of objects are approximate. Other assistants have their own limits; check the one you use.

One visit worked through: messy notes in, a clear report out

This is a progress visit to a two-storey rear extension, with the ground-floor walls up and the first-floor structure half fixed. The notes are the kind most of us take: abbreviated, uncertain in places, typed on a phone. Everything here is illustrative.

Visit 9, Thu 29 Oct, 10:15-11:30. Dry, cold. Present: contractor's site manager (SM), me.
Client not present. Purpose: progress, first floor structure.
P1 wide from garden, rear elevation. Ground floor blockwork complete to first floor level.
Steel beam over rear opening in place, bearing at both ends looks fine?? (P2, P3).
Engineer's drawing not on site - asked.
First floor joists going in, about half done. Spacing looks wider than drawing? didn't measure. P4
Told SM to stop fixing and check spacing against engineer's drawing first. He agreed.
Window opening W2 - lintel in, width looks ok. P5
Brickwork N elevation: joints uneven left of W2, P6. SM says bricklayer will rake out and repoint.
Membrane at slab edge: couldn't see, covered by blocks already.
Scaffold: tie at top lift missing near rear corner? P7. Told SM.
Didn't go up to roof. Drainage trench backfilled, not seen.
Skip bins fine.
Need from contractor: engineer's drawing on site, window delivery date, updated programme.

Put those notes through the prompt below, correct and add to the result, and the report table reads like this:

Below the table, the report carries a "Not observed or not accessible" list: roof structure (not accessed), drainage trench (backfilled, not seen) and slab-edge membrane (covered by blockwork, so not seen). The four-line summary at the top then says what the visit found, what was raised on site, what could not be checked and what is outstanding. Where a row's action is the architect's own addition, such as rechecking the repointing, it is fine for it to be there; the rule is that the assistant may not invent one.

Row 7 is deliberate. It records contractor deliverables as actions, not as observations, so nobody can read the report as saying the programme was seen. Count what has happened to the raw notes: 14 lines became seven table rows, a four-line summary, a three-item "not observed" list and seven photo captions. Everything in the output traces back to the input, which is the test.

A prompt with seven rules

This prompt produces the table above. Run it in a business-plan assistant where your content is not used for training by default, and attach the photos in batches.

You are helping an architect write a site visit report. Use ONLY the notes and
photo list below.

Rules:
1. Every item must come from my notes. Do not add items, comments or conclusions.
2. Keep my level of certainty. If I wrote "looks", "appears" or "?", write
   "appeared" and say it was not measured or verified. Never upgrade to
   "complies", "satisfactory", "correct" or "checked" unless I used that word
   about that item.
3. Use only these statuses: Observed, Instructed, Reported, Not observed,
   Not accessible. Anything I could not see goes under Not observed.
4. Instruction column: only what I told someone on site, and who. Otherwise
   "None given".
5. Action column: who does what by when, only if my notes say so. If my notes
   give an action without an owner or date, write "to be agreed" for the
   missing part. If there is no action, write "None".
6. Do not cite standards, regulations or drawings that are not named in my notes.
7. If anything in my notes is ambiguous, list it under "Questions for the
   architect" instead of guessing.

Output: (a) a four-line summary; (b) a table with Ref | Location | Observation |
Instruction | Action | Photos; (c) Not observed or not accessible; (d) Questions
for the architect; (e) a caption for every photo, 25 words maximum, describing
only what is visible plus my note.

NOTES:
[paste]
PHOTO LIST:
[P1 = ..., P2 = ...]

Rule 7 is the quiet hero. An assistant that is allowed to ask will return questions instead of making up an answer. For this visit the list might read:

1. The beam note says the bearing "looks fine??". Was the bearing length measured, or judged by eye?
2. The W2 note says the lintel width "looks ok". Against which drawing?
3. The scaffold note ends "Told SM". Is that an instruction to the contractor or a mention in passing?

You answer in two minutes: by eye from the ground; against the architectural elevation, not the engineer's drawing; a mention in passing. Those answers are exactly why rows 2, 4 and 6 above carry "not measured", "appeared" and "site manager told" rather than anything firmer. Without rule 7, the assistant would have picked an answer for you.

Where the first draft claimed more than the notes

Even with those rules, expect slips. This is what a first draft of this visit tends to contain, set against what the notes support.

Each correction is the same move: shrink the claim back to the note. The most dangerous line is the membrane one. The note said "couldn't see" and the draft said "inspected". If that sentence had gone out, the report would have recorded a check that never happened, on a detail that is now permanently hidden.

The same method at foundation stage, where the pour hides everything

Earlier stages make the "not observed" list more important, because concrete will cover what you saw or failed to see. A second illustration shows how the five status words handle a visit the day before a pour. The notes: "Trench excavated, P1. Reinforcement visible, P2, spacing not measured. Bottom wet in one corner, P3. Told SM to pump out and call me before pouring. Pour booked Friday."

The report items that survive the rules are: Observed, trench excavated; Observed, reinforcement present, spacing and cover not measured; Observed, standing water in one corner of the excavation; Instructed, site manager asked to remove the water and contact the architect before any concrete is placed; Not observed, formwork dimensions and the pour itself, which is scheduled for Friday. A careless draft writes "reinforcement checked" and "pour scheduled as planned". The first is a claim you did not make, and the second turns a booking into an event. If you want a check against a specific drawing, say so in the notes and name the drawing, otherwise the report cannot imply one.

Photo captions that earn their place

A caption of "Beam" is useless in a dispute two years later. A good caption has four parts: the photo number, the location, what is visible, and why it matters, with a link to the report item. Compare these:

• Weak: "Steel beam."
• Assistant's first attempt: "Steel beam installed over the rear opening with full bearing at both ends."
• Corrected: "P2: Steel beam over the rear opening, bearing at one end, seen from ground level. Bearing length not measured (item 2)."

The assistant's attempt is where the trouble sits: it describes bearing the photo cannot show. Let it propose captions, then check each against what the picture actually shows and what you wrote. Do not ask it to count items in a photo (fixings, bricks, joists), and do not rely on it for dimensions or scale; Anthropic's own documentation describes its counts and its location estimates as approximate. Crop out people, vehicle plates and neighbouring windows before uploading.

For a structural engineer's version of the same job, with a different emphasis on defects and limits, see the inspection report tutorial for photos and notes.

Instruction and action columns without overstepping

The instruction column records what you told someone on site, and who. It is not a formal instruction under the contract. If your contract requires a written instruction to change cost or time, the report is not that instrument, and the report should say so in its standing notes. Where something is urgent for safety, phone the contractor and confirm by email the same day; do not wait for the report to go out.

The action column names an owner and a date. If your notes do not give a date, the assistant must write "to be agreed", not pick one. A contractor who reads "by Friday" in a report you issued will take it as a commitment, so make sure you are the one who set it.

Carrying open items from one visit to the next

A report is more useful when it shows what changed since last time. Keep an open-items list from the previous visit and give it to the assistant with the new notes, with one strict instruction: an item is closed only if the new notes say it was seen resolved. Otherwise it stays open, with the evidence of what was said.

The last row is the one to watch. Silence is not closure. Assistants love to tidy a list by marking everything not mentioned as resolved, so the prompt must say that "no new information" is a valid answer.

A plain-English summary for the client

Clients rarely want a six-column table. Once the report is checked, the same assistant can write a short summary for the client from the table alone, with one rule that is easy to forget: no reassurance. Ask for 90 words, no technical terms, nothing described as fine, and nothing that is not in the table.

Work is progressing. The ground-floor walls are complete and about half of the first-floor joists are fixed. Two points were raised with the contractor on site and are being followed up: the spacing of the joists, and a scaffold tie. Some parts could not be seen on this visit, so this report says nothing about them. The contractor owes us three pieces of information before the next visit.

That is illustrative output. Check it as you would the report: every sentence must trace to a row. The first draft will often add a closing line such as "overall the project is on track", which the table does not support. Delete it.

Wording for the limits of a visit

Every report should say what a visit is and is not. A standing paragraph, adapted to your appointment, keeps expectations honest:

This report records observations made during a visual visit on [date]. Areas that were not accessible, were covered or were not yet started are listed under "Not observed". It is not a full inspection or certification of the works, and no testing or opening up was carried out. Matters reported by others are marked as reported.

Treat that as a starting draft, not as approved wording. What your appointment promises, and what your insurer expects a report to say, are matters for your own terms and adviser; the tutorial on what to tell your insurer about AI is a good prompt for that conversation.

Minutes per visit, and the checking time that stays

A visit that used to be a 75-minute walk-round followed by 90 minutes of typing at the desk takes 165 minutes in total. With the method above it is 75 minutes on site, 15 minutes tidying notes and numbering photos, 10 minutes running the prompt and 25 minutes checking: 125 minutes. That is a saving of about 40 minutes a visit, or about 160 minutes a month on four visits. These figures are an illustration, not a benchmark.

The honest part: the checking time does not shrink as the assistant improves, and it should not. What you gain is consistency (every report has the same columns and the same "not observed" list), fewer omissions, and notes that are better because you now write them knowing exactly where each word will end up. If you also minute client meetings, the same discipline applies in the same-day meeting minutes workflow.

A seven-point check before the report goes out

1. Every table row traces to a line in your notes or a photo. Tick them off one by one.
2. Read every observation for upgraded words: complies, correct, satisfactory, checked, verified, approved. Each must match a word you used.
3. The "Not observed" list contains everything you did not see, including areas on the schedule that you skipped.
4. Instructions match what you said, to the right person, and nothing was added.
5. Actions have owners and dates that you actually set.
6. Photo numbers match the images, and each caption describes only what is visible plus your note.
7. No standards, regulations or measurements appear that were not in your notes.

Keep the issued report, your original notes and the photos together. The AI chat is a drafting tool, not part of the record; if the report is ever questioned, the notes and photos are what show what you saw.

Questions about AI and site visit reports

What if the contractor disputes a line in my report?

Go back to your notes and photos. If the line matches them, reply with the evidence. If the contractor adds information, record it in the next report as reported, with the date and who said it. Never edit an issued report quietly; issue a revised version, keep both, and note what changed and why.

What if I forgot to number the photos on site?

Number them at your desk in camera-roll order before uploading, and add a location to your notes only where you are certain. For any photo you cannot place, write location unconfirmed and keep it out of the table. Do not let the assistant guess locations from what it sees, because similar elevations and rooms look alike.

Can I use the same method for snagging or final walk-rounds?

Yes, with a stricter structure. Number each room or area, record every defect as its own item with a photo, and keep a long Not observed list for anything you could not open, test or reach. The assistant still only rephrases what you saw, and a status column on a snag list is one only you should change.

Further reads

• AI for Electrical Project Management: Diaries, RFIs, Snag Lists — Site diaries and RFIs, the paperwork that runs beside visit reports.
• AI Software for Architectural Drawings: Notes, Schedules and Sheet Checks — Checking drawing notes and schedules before they reach site.
• AI Policy for Architects: A Ready-to-Use Checklist and Clauses — Write down which AI tools may touch site photos and reports.
• 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 Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Anthropic's vision documentation (image limits, metadata handling and accuracy caveats), checked October 2026. The report structure is a practical method written for this tutorial.