From Screenshots to a Spreadsheet: Can Phone Automation Do the Typing?

Around the sixth page, it drifts

Someone I know does product sourcing. Every morning he transcribes three things off his phone: competitor prices, ranking positions, and his own sales figures.

Three to five rows per page, five or six pages. It does not sound like much.

Then one day he compared two days of figures side by side and found a number that differed by a factor of ten. That was the day he slipped on page four.

His summary was practical: “I know it drifts when I go long. I just don’t know which page it happens on.”

That is the nature of this work: not whether you will make a mistake, but which page you make it on.

---

Three reasons it suits automation

Not everything should be handed to a machine. Data collection is one of the few that meets all the conditions:

• High frequency: daily, sometimes several times a day.
• Fixed steps: open the page, read the figures, note them down. Same action on every page.
• Verifiable result: a wrong number is obvious at a glance.

The third matters most, because it means you can cheaply check whether the machine is doing it right. Spot-check the first few runs, confirm the accuracy, and you can relax afterwards.

Compare it with another job: deciding whether a lead is worth pursuing. High frequency, but the steps are not fixed and the result is not checkable, so it does not qualify.

---

Three stages from screen to sheet

Taken apart, there are three stages, each with its own trap.

Stage one: read it

Two ways.

• Read the page text where the page carries structure. This is the most accurate route.
• Recognise text in an image where only a picture exists, such as figures baked into a chart.

A good approach is to mix them: read text where the field has it, recognise the image where it does not. There is no need to force one method everywhere.

Stage two: structure it

The text that comes out has to become meaningful rows and columns.

The key here is defining the fields first: this column is the date, this one the product name, this one the price. Without that, what you have is a lump of text that is useless three months later.

Once the fields are set, define the mapping — which position on the page corresponds to which column. The most common failure here is the data order changing: one day a field moves, the machine keeps reading the old position, and the columns shift. So read by field name rather than position wherever possible.

Stage three: write it

The last step puts the result into the sheet.

Appending to one fixed sheet is the recommended approach: one new row per run rather than overwriting. Three months later you have data you can chart instead of daily snapshots.

This step is the one most often skipped, and skipping it wastes the first two. The value of data collection is not what you read today; it is what thirty days of reading shows.

---

Two safeguards

Both are cheap to add.

• A record at each stage. Which page you reached, how many fields that page yielded, whether anything failed to read. With that record, a failure can be traced to a stage rather than just reported as “the result looks wrong”.
• A stop on anomalous values. If a reading is wildly different from the last one — the price suddenly halved, for instance — do not write it in. Screenshot it and wait for confirmation.

The second guard is against acting on a misread. A misread is not the problem; a wrong number written into the sheet and then acted on is.

---

How to start

Get one page working.

Read it, structure it, write it, then confirm three things: is the reading accurate, are the fields right, and does it land in the right place. Spot-check the first few runs before moving on.

Add the second page once the first runs steadily. Starting with five pages leaves you unable to tell whether the fault is in the reading, the fields, or the placement — and troubleshooting costs more than doing it one at a time would have.

In terms of time, an hour or two gets one page reading and writing. The rest goes into verification.

---

The value of an output

One last point that gets overlooked.

Most tools only solve how to operate — where to tap, where to type. How the data becomes something readable afterwards is rarely discussed.

In practice, “read it” and “can use it” are separated by that assembling step. The same act of reading the screen, with an output attached, turns from a record into something that supports decisions.

That is why this work deserves its own treatment: what it saves is not just the transcribing, but the second pass of assembling that used to follow it.

The software is completely free and runs on your own computer. For a related idea, read letting AI look first, then decide; for keeping an eye on competitors daily, see tracking competitor prices and content.