Images, Drawings & Documents

Multimodal simply means Claude accepts more than text - you can attach images (photos, screenshots, sketches, marked-up drawings, charts) and documents (PDFs, including many scanned ones read via vision, plus other file types), and Claude reasons over them alongside your words. You upload the file, ask your question in plain language, and it responds as if it had looked at what you looked at. No special format is needed; a phone photo or a screenshot is fine.

What it reads genuinely well is the interpretive, qualitative layer. Give it a photo of a room and it will describe the materials and finishes, the quality of light, the style, what looks worn or cluttered - useful for a condition note, a moodboard description, or a first read of a space. Give it a hand sketch and it will infer the layout and intent and discuss it with you, which makes it a fast sounding-board for a rough idea. Give it a marked-up drawing and it will read your annotations and the general arrangement and help you turn them into a written instruction or a snag list. Give it text in a PDF - a brief, a spec, a product data sheet, a byelaw chapter, a scanned report - and it will summarise it, pull out the clauses or figures that matter, compare it against another document, or answer questions across many pages at once, thanks to its large context window. This last use alone - turning a 60-page PDF into the three paragraphs that bind your site - saves hours a week.

The common thread is that Claude is excellent at the meaning in a visual input and unreliable at the metrics. It grasps "this is an open-plan living-dining with a kitchen along the north wall and the client has scribbled move the island" far better than it grasps "the island is 1,800 mm long." Lean on it for the former; never trust it for the latter.

It helps to picture what actually reaches Claude. A photo or a PDF page arrives as a flat grid of coloured dots - an image - and Claude interprets that image the way a knowledgeable person glancing at a print would: it recognises shapes, words it can make out, the overall arrangement. What it does not receive is the drawing's real geometry - the coordinates, the true lengths, the layers of a CAD file - because none of that survives in a picture of the drawing. So the very same input that lets it say "this looks like a two-bedroom flat with a long north-facing living room" gives it no reliable way to say how long that room is. Holding that mental model - it sees a picture, not a model - explains every strength and every failure in this lesson.

Here is the wall, stated as plainly as it can be: Claude cannot measure precisely from an image, and it cannot read fine CAD linework reliably. Ask it for the dimension of a room off a plan, the area of a space, the number of parking bays on a layout, or the exact reading on a dimension line, and it may give you a crisp number - but that number is an estimate dressed as a fact, and it is often wrong. It is not measuring; it is guessing at what a plausible answer would look like, and because nothing in the machine flags uncertainty, the wrong number arrives with exactly the same confidence as a right one. This is the image-based face of hallucination, and on drawings it is dangerous precisely because the output looks authoritative.

Why does this happen? An image reaches Claude as pixels it interprets, not as the underlying CAD geometry with real coordinates. Fine lines blur together, small text is missed or misread, a dimension string can be transcribed with a digit wrong, symbols get confused, and anything at small scale - a stair going the wrong way, a door swing, a hatch pattern - is easily misread. A human draughtsperson checks a suspicious number against the scale and the context; Claude has no scale rule and no site sense, so it fills the gap with plausibility. The failure is quiet, which is what makes it a liability: it does not say "I can't read this"; it says "3,600 mm" and moves on.

So the rule is absolute and worth memorising: never trust a dimension, area, count or quantity that Claude reads off an image - verify every one against the actual drawing, the model, or a measured survey. Use Claude to discuss a drawing, describe it, draft notes about it, or explain a detail - the qualitative work it is good at. For anything numeric or safety-related read from an image, treat its output as a prompt to go and check the real source, never as the source itself. This is the same scrutiny-scaled-to-stakes discipline from Lesson 0.1, at its sharpest: a drawing dimension is about as high-stakes as reading gets.

Within those limits there is a lot of good, everyday practice. A few habits get the most from multimodal Claude while staying safe.

Give clear images and one clear ask. A sharp, well-lit photo or a clean scan reads far better than a dark, skewed phone snap. Tell Claude what to look at and what you want - "describe the condition of the joinery in this photo for a survey note" beats "what do you see." For a PDF, point it at the part you care about: "in this byelaw chapter, summarise only the clauses about setbacks and height, and quote the clause numbers so I can check them." Asking it to quote clause numbers or locations is a small trick with a big payoff - it gives you an anchor to verify against, and it keeps Claude tethered to the document rather than drifting into its training data.

Supply the numbers instead of asking Claude to read them. This is the key move that turns the hard limit from a blocker into a non-issue. Rather than "what's the area of this room?", tell Claude the dimensions you have measured and let it do the arithmetic and reasoning: "this room is 4.2 m by 3.6 m - is that enough for a king bed with 700 mm circulation both sides?" Now Claude is doing what it is good at (reasoning from given facts) and you have kept control of the one thing it is bad at (reading the measurement). Even then, sanity-check its maths, as Module 7 stresses.

Combine visual input with everything else you have learned. Feed a marked-up drawing plus your written intent and ask for a snag list as an Artifact. Load a product PDF into a client Project so every chat can reference it. Use a site photo to seed a condition report you then correct. And keep confidentiality in view - a client's drawings and site photos are sensitive documents; mind your plan's data settings before uploading, and anonymise where you can. Used this way, Claude's eyes become a real extension of your own - as long as you never hand it the ruler.

Say a client sends a PDF of an existing flat's floor plan and asks whether their furniture will fit. The unsafe way is to upload it and ask "measure the living room and tell me if a 3-seater sofa fits." Claude will answer with confident numbers it read off the image - and you would be staking advice on a guess. The safe way uses everything in this module:

I have uploaded a floor plan PDF. Do NOT read any dimensions off the
drawing - you cannot measure it reliably and I do not want guessed
numbers.

Instead: (1) describe the layout and name each room as labelled, and
flag anything ambiguous; (2) list the questions I should answer with a
tape measure before advising on furniture (which walls, which openings,
circulation); (3) once I give you measured dimensions, help me reason
about fit.

Claude describes the plan (good - qualitative), tells you exactly what to go and measure (good - it turns its own weakness into your checklist), and waits for real numbers. You measure on site, come back with "living room 5.1 m by 3.8 m, one 1.2 m opening on the east wall," and now Claude reasons genuinely usefully about sofa size and circulation - because you supplied the metrics.

That pattern - let Claude interpret, make yourself supply the measurements, verify anything numeric at the real source - is the whole lesson in one workflow. It keeps the superpower (a tireless colleague who can read your photos, sketches and PDFs and turn a document mountain into a summary) and defuses the landmine (a confident-wrong number off a drawing). Master it and Claude becomes a daily part of how you handle visual information, without ever once being mistaken for a scale rule, a measured survey, or a set of coordinated CAD.