# Working from your own files and images

> Hand an assistant a screenshot of a table instead of describing it, get the rows back for a spreadsheet, and count the cells that came back wrong. Then compare two versions of a contract with a brief that says which differences matter.

Page: https://schubergphilis.github.io/ai-training/using-agents/working-from-files/
License: CC BY-SA 4.0, https://creativecommons.org/licenses/by-sa/4.0/

In this lesson we hand an assistant a screenshot of a table and get the
rows back for a spreadsheet. First we try to describe the table in words
and see where that goes wrong. Then we upload the picture with a brief
that says what to keep exact. The check is a count of the cells that came
back wrong. A second run compares two versions of a short contract, where the
brief says which differences matter. By the end you have a rule for
working with files and images: when to upload and when to type.

You need a chat assistant that accepts images. Most do, and the one this
lesson checks its limits against takes pictures in the common formats up
to 8000 by 8000 pixels each, about twenty of them in one message, and a
file size limit that changes between products and over time, so check the
one your product states (Vision, Claude Platform documentation). Every response shown here is illustrative. The
author wrote it to show the pattern, and your model gives a different
text.

## The table

Here is the material. It is a stock count from Brindlewick Instruments,
the fictional company from the previous lesson, and it exists only as a
picture: someone photographed the count sheet, or took a screenshot of a
system you have no export from. The task is to get these six rows into a
spreadsheet.

![A screenshot of a table titled "Storeroom B, stock count, 18 September" with six rows and the columns Item code, Description, Shelf, Counted, Expected and Difference.](https://schubergphilis.github.io/ai-training/using-agents/working-from-files/stock-count.png)

Look at it once now. Thirty-six cells, some of them one digit, some
of them a minus sign in front of a digit. You could type them, and typing
is where a `-1` becomes a `1`. The point of this lesson is to make the
assistant do the typing and to make the check quick.

## Describe the table in words

The first instinct is to describe what you see. Before you read on,
predict what comes back when you do that. Where does the assistant get the
values you leave out?

#### Prompt (illustrative, not a recorded transcript)

```text
I have a stock count table with six rows. The columns are item code,
description, shelf, counted, expected and difference. The items are a
torque wrench, a calibration weight, two digital calipers, a thermocouple
probe and a spirit level. The item codes start with BI and a dash and a
four-digit number. Can you give me the rows to paste into a spreadsheet?
```

#### Response

````text
Here is the table as tab-separated rows. I have filled in the codes and
numbers as placeholders where you did not give them, so replace those with
the real values.

```text
Item code	Description	Shelf	Counted	Expected	Difference
BI-0001	Torque wrench	B1	10	10	0
BI-0002	Calibration weight	B1	5	5	0
BI-0003	Digital caliper	B2	12	12	0
BI-0004	Digital caliper	B2	12	12	0
BI-0005	Thermocouple probe	B3	8	8	0
BI-0006	Spirit level	B3	4	4	0
```
````

The response says so itself, and it is still no use. The numbers are
made up, because the prompt gave none, and the two calipers are
indistinguishable because the prompt did not say how they differ. You
would now type thirty numbers into the assistant's placeholders, which is
the work you wanted to hand over. A description is a lossy copy of the
picture, and the loss is in the exact values, which is the one part of
the task that matters.

## Upload the screenshot

Now attach the picture. Save the screenshot from this page to your
computer, open a fresh chat, add the file to the message, and write the
ask under it. The brief has the parts you used for the expense guideline
in the previous lesson. The goal is rows for a spreadsheet. The context is
the attached file, named as the only source. The limits say what to keep
exact and what to leave out. The done-criterion is one you can check
against the picture.

#### Prompt (illustrative, not a recorded transcript)

```text
[stock-count.png attached]

Attached is a screenshot of a stock count. Give me every row as
tab-separated values that I can paste into a spreadsheet, with the header
row first and the rows in the same order as the picture. Keep the item
codes and the numbers exactly as shown, including the sign. Do not add a
total or any row that is not in the picture. Where a cell is unreadable,
write a question mark instead of guessing.
```

#### Response

````text
```text
Item code	Description	Shelf	Counted	Expected	Difference
BI-1042	Torque wrench 20-100 Nm	B3	14	15	-1
BI-1108	Calibration weight 1 kg	B1	8	8	0
BI-2260	Digital caliper 150 mm	B4	21	19	+2
BI-2271	Digital caliper 300 mm	B4	6	6	0
BI-3015	Thermocouple probe, type K	B2	37	40	-3
BI-3390	Spirit level 600 mm	B5	11	11	0
```
````

Paste the rows into a spreadsheet and put the picture next to it. Go
through the cells one at a time and count the ones that differ from the
picture. On a clear screenshot like this one the count is zero or one,
and the cells that go wrong are the ones a person mistypes too: a sign
dropped from a negative number, or a `1` read as a `7`. The vendor's own
documentation says that the model reads small, blurry or rotated text
badly and that its counts of objects are approximate. It recommends text
that is legible in the picture. A crop that makes the text larger should
still show the context around it, such as the header row (Vision, Claude Platform documentation). The check is what tells you which kind of picture you have.

The upload changed the task. You wrote one sentence about the file and
three about what to keep exact, and the assistant did the thirty-six
cells. The same brief works for a photo of a whiteboard, a PDF price
list, a spreadsheet export, or a scanned form. Name the attachment as the
source and say what must be copied exactly. Give the model a way to say
"I can't read this", so that it doesn't fill the gap.

#### Checkpoint: Which ask gets the table?

You have the screenshot of the stock count and a chat open. Which ask gets
you rows you can paste and check in a minute?

- I have a stock count with six rows and six columns, items like a torque wrench and two calipers. Give me the rows for a spreadsheet, codes included.
- What is in this image?
- Attached is a screenshot of a stock count. Give me every row as tab-separated values, header first, codes and numbers exactly as shown. Write a question mark for any cell you cannot read.
- Type up the attached table for me and make sure it is accurate.

## Compare two versions of a contract

Here is a second use of the same move. A short service agreement has
come back from the other side with changes, and you want to know what
changed. The
versions are below, and each is short on purpose so that you can check
the answer. In your own work they would be two files, and you attach both.

> **Service agreement, calibration services (version 1)**
>
> 1. Brindlewick Instruments calibrates the Client's measuring equipment
>    listed in Schedule A once every twelve months.
> 2. The fee is 1,200 euros per visit, invoiced after each visit and
>    payable within 30 days.
> 3. Either party may end this agreement with 60 days' written notice.
> 4. Brindlewick repeats a faulty calibration free of charge when the
>    Client reports it within 14 days of the visit.
> 5. Brindlewick keeps the calibration records for five years and provides
>    copies on request.
> 6. Brindlewick may not subcontract the work without the Client's written
>    consent.

> **Service agreement, calibration services (version 2)**
>
> 1. Brindlewick Instruments calibrates the Client's measuring equipment
>    listed in Schedule A once a year.
> 2. The fee is 1,350 euros per visit, invoiced after each visit and
>    payable within 30 days.
> 3. Either party may end this agreement with 30 days' written notice.
> 4. Brindlewick repeats a faulty calibration without charge when the Client
>    reports it within 14 days of the visit.
> 5. Brindlewick keeps the calibration records for five years and provides
>    copies on request.
> 6. Brindlewick may subcontract the work to a certified laboratory and
>    remains responsible for it.

"Compare these two" gives you a list in which a fee increase and a changed
turn of phrase sit side by side with the same weight. The brief says which
differences matter and what to do with the rest.

#### Prompt (illustrative, not a recorded transcript)

```text
[contract-v1.txt and contract-v2.txt attached]

Attached are two versions of a service agreement. List every difference
that changes an amount, a period, a deadline or an obligation. For each
one, quote the sentence from both versions. Rewording that keeps the
meaning goes in one line at the end, and if a clause is unchanged, do not
mention it.
```

#### Response

```text
1. **Fee.** Version 1: "The fee is 1,200 euros per visit". Version 2:
   "The fee is 1,350 euros per visit".
2. **Notice period.** Version 1: "with 60 days' written notice". Version
   2: "with 30 days' written notice".
3. **Subcontracting.** Version 1: "Brindlewick may not subcontract the
   work without the Client's written consent." Version 2: "Brindlewick may
   subcontract the work to a certified laboratory and remains responsible
   for it." The Client's consent is no longer needed.

Rewording only: clause 1 says "once a year" instead of "once every twelve
months", and clause 4 says "without charge" instead of "free of charge".
```

Check it against the two texts. Each of the three items quotes both
versions, so the check is finding two sentences, and the reworded clauses
are named and set aside rather than mixed in. The brief did that. It named
the kinds of change that matter (amount, period, deadline, obligation),
asked for the quote behind each, and gave the rewording a place to go.

#### Checkpoint: Which differences does the brief ask for?

Buckets:

- A change the brief asks for
- Rewording the brief sets aside

Items:

- The fee goes from 1,200 to 1,350 euros per visit
- "Once every twelve months" becomes "once a year"
- The notice period goes from 60 to 30 days
- "Free of charge" becomes "without charge"
- Subcontracting no longer needs the Client's written consent

## What an upload discloses

An attachment goes to the same place as pasted text, under the same
terms, and the model reads all of it. The difference is that you read
pasted text as you paste it, and you rarely read a file before you attach
it. In the lesson on redacting before you paste you read a document with
the task in mind and split it into what the task uses and what says who is
involved. Do the same with a file, and where a file holds both, make a
copy that holds only the first part: crop the screenshot, or export the
one sheet.

#### Pitfall: A screenshot of the whole window

A colleague wants help with the formula in one column of a spreadsheet and
takes a screenshot of the whole window. The picture shows the column, and
also the browser tabs behind it, a customer's name in a chat notification
in the corner, and the salary column next to the one they asked about. The
assistant answers the formula question well, and the whole picture went
with it. An attachment discloses its whole content, the same as pasted
text. Crop to the range the task needs before you attach.

#### Checkpoint: Which file goes in the brief?

You want a chart of the daily totals column from the team's spreadsheet,
two hundred rows of it. The sheet also has a column of names and a column
of salaries. What do you attach?

- Attach the whole spreadsheet and ask for a chart of the one column.
- Copy the one column into a new file, attach that, and name it as the source in the brief.
- Type the two hundred numbers into the chat instead of attaching anything.

## When to upload and when to type

Upload when the material is long, structured or visual: a table, a
diagram, a photo of a whiteboard, a document with numbering you want kept.
Describing it loses the exact values, and the exact values are the task.
Type when the material is one or two sentences, or when the file holds
things the task doesn't need and you can't cut them out in a minute.

Either way the brief is the one you wrote for the expense guideline in
the previous lesson, with the attachment in place of the pasted text. Name the file as the source, say which values must be copied exactly, give the model a way to say it can't read something, and
check the result against the file. A read of a picture is a transcription,
and a transcription error is silent until you look.

## Exercise

Save the stock count screenshot from this page, open a fresh chat, attach
it, and send the brief from the "Upload the screenshot" section. Paste the
rows into a spreadsheet and compare them with the picture, cell by cell,
and write down the number of cells that differ and what each one says. It
takes about five minutes and produces one number and at most a few lines.
Doing it once by hand shows you what kind of error a picture read makes,
so you know what to check for next time. A good result is a count you can
name, with a line for each wrong cell, and zero is a fine count. Which of
the wrong cells, if any, would you have got wrong yourself if you had
typed the table?

Stretch: Shrink the screenshot to half its width in any image editor, run the same brief on the smaller copy, and see whether the count of wrong cells changes.

## Recap

1. A description of a table is a lossy copy of it, and the loss is in the exact values. Attach the picture or the file, and let the model do the typing.
2. The brief for an upload has the same four parts as any other. Name the attachment as the only source, say which values must be copied exactly, and give the model a way to say it can't read a cell.
3. Check the result against the file, cell by cell. A model reads small, blurry or rotated text badly, and a transcription error is silent until you look (Vision, Claude Platform documentation).
4. When comparing two files, say which kinds of difference matter and where the rest should go, and ask for the quote from both versions behind each one.
5. An attachment discloses its whole content, the same as pasted text. Crop or copy out the part the task needs before you attach.

## References

- Vision, Claude Platform documentation, https://platform.claude.com/docs/en/build-with-claude/vision