Lesson 30 of 32
The Senses of the Machine
Files, Images & Voice
These assistants can now look at a photograph, read a document, work through a sheet of figures and listen to your voice - so stop labouring to describe what you could simply show. But a picture without a purpose is a ship without a port: name what the thing is, supply the context the camera could not capture, point to the exact page or column that matters, and say what you want out of it.
Claude's senses are strongest on the scroll. Long PDFs, several documents at once, a spreadsheet and the circular that governs it - it will hold all of them and answer across them, which is why the pointed-page technique in this chapter pays off here more than anywhere. Put the material first and the instruction last, inside tags: <document>, <scope>, <task>. For files you come back to - a policy manual, a syllabus, last year's rate card - put them in a Project instead of attaching them again in every chat. It reads photographs too, and reads the layout of a scanned or printed page unusually well, which matters when the thing you are holding up is a bill or a form rather than a leaf.
From the book Attach PDFs, documents, spreadsheets or photos with the paper-clip or plus button; Claude handles very long documents well, so you can ask about a whole report and request page numbers. Put files you use again and again (a policy manual, a syllabus) into a Project so every chat can refer to them.
- Material first, instruction last - in <document>, <scope> and <task> tags.
- Say “using only the attached file” and ask for page numbers.
- Ask for the exact quotation, not a paraphrase, where it matters.
- Keep files you reuse in a Project rather than re-attaching them.
- For spoken back-and-forth in an Indian language, use another tool.
Instead of
What is wrong with this leaf?
Claude version
<what_i_am_showing_you> Two photographs from my tomato field: one leaf from an affected plant, and the same plant from a few steps back so you can see how much of it is involved. </what_i_am_showing_you> <context_the_camera_cannot_see> Near Kolar, Karnataka. The monsoon has just set in and the nights are humid. Roughly 20 plants of 300 are affected, mostly older lower leaves. It has spread over about a week. I have sprayed nothing so far. </context_the_camera_cannot_see> <if_unsure> Say plainly when the photograph cannot separate two causes, and name the picture that would settle it - the underside, the stem at soil level, a cut through the stem. Do not pick the commonest disease merely because something has to be named. </if_unsure> <task> Work in three separate parts, in this order. First, observation alone: describe what is in the image, with no diagnosis at all. Second, reasoning: think aloud about which causes fit what you just described and which do not, and why. Third, the shortlist: two or three causes, each with the field check that distinguishes it, the simple first step, and the question I should put to the Krishi Vigyan Kendra. </task>
Then loop it
- Here is the underside of the leaf. Go back to your second part and tell me what has changed in the reasoning - not just a new conclusion.
- Now write the questions for the agriculture officer in Kannada, short enough to read aloud on a phone, with the English beneath each one.
Why it worksThe photo shows the problem, but the words supply what the camera cannot - place, season, spread - and ask the AI to admit when it cannot see enough.
Instead of
Can you summarise this PDF for me?
Claude version
<document> [Attach the 40-page pension circular here, above the instruction that follows.] </document> <scope> This circular is the only source. Nothing from your general knowledge of pension rules, and nothing from any other circular you may know. Section 4 and the annexures are the heart of it. Read the remainder only to find anything that modifies Section 4, and tell me where that sits. </scope> <what_i_need> The sequence an employee passes through, from first application to final sanction. Each document that must be produced, tied to the stage at which it is demanded. Every time limit the circular imposes, on the employee or on the office, wherever it appears. </what_i_need> <verification_rule> Give the page in brackets after each point. Where a point rests on a sentence I may be asked to defend at my desk, quote that sentence rather than paraphrasing it. Where the circular says nothing, write “Not stated”. Where two parts conflict, write “Conflict - pages X and Y”. Do not smooth either one over. </verification_rule> <task> Start by telling me which pages read cleanly and which arrived as poor scans. Then produce the three lists above. I am a junior clerk and employees will act on what I tell them, so a gap you have flagged is useful to me and a confident wrong answer is not. </task>
Then loop it
- Page 18 mentions a form. Quote that paragraph word for word, and tell me whether the wording is “shall” or “may”.
- Your step 5 and the annexure on page 33 conflict over the signing authority. Quote both passages and say which the circular itself gives effect to.
- Now make an Artifact: a one-page checklist table - Step | Document | Time limit | Page - for the employees who come to my desk. We will correct it in place as I check it against the circular.
Why it works"Using only the attached PDF" anchors the AI to the document, and page numbers let Fatima verify every answer herself.
Instead of
(spoken) Bill… why so much?
Claude version
(An honest note first: for a spoken back-and-forth in Tamil, ChatGPT's Voice mode or Gemini Live is the better place. Claude is the better place for reading the bill itself - so either have a grandchild type this, or dictate it with your phone keyboard's own microphone key.) <what_i_am_showing_you> A photograph of this month's electricity bill, and one of last month's. I have covered the consumer number and the address with my thumb in both. </what_i_am_showing_you> <about_me> I am 68, in Madurai, and I read very little. Someone will read your answer out to me, so write it to be read aloud: simple Tamil, short sentences, English underneath. </about_me> <what_to_tell_me> One: the amount to pay, and the last date, each on its own line with nothing else beside them. Two: the units used this month against the units last month. Three: in two or three sentences, why the bill is higher. Four: one or two ways to use less current that cost nothing. </what_to_tell_me> <verification_rule> Quote each figure exactly as it is printed on the bill before you use it in a sentence. If a figure is unclear in the photograph, name it and ask for that corner again - do not read it as the number it most resembles. </verification_rule> <task> Answer in that order, with nothing at all before the amount and the date. </task>
Then loop it
- Read the last date back to me as a date and as a day of the week, so I can be sure you read that figure and not the one next to it.
- Is there a late fee or any extra charge on this bill? Yes or no first, then quote the line it comes from.
Why it worksVoice removes the barrier of typing and reading, but the prompt still states the aim, the language and the format - and personal details are hidden first.
Loop it
When you share a photo or file, the first answer tells you what the AI actually saw. Check that first: did it read the right page, the right column, the right figure? If it misread something, correct it - "The amount in row 7 is ₹1,250, not ₹125" - and ask it to redo the answer. If a photo was blurred, send a clearer one or a close-up. Then narrow the focus, ask for page numbers or quotes to verify, and turn the result into something you can reuse: a checklist, a table or a spoken summary.