When the link comes back wrong

Your answer is written in your own ChatGPT or Claude window. We never see it, which is the point, and also means we cannot quietly correct it on the way. So it is worth knowing what a good one looks like and what the three broken ones look like.

How it is supposed to go

  1. You answer a few questions here and we write the prompt.
  2. You paste it into ChatGPT or Claude and it goes to work.
  3. It writes a table, and then after the table it writes a link. Pressing that link is what brings the list here.

The commonest problem is not breakage at all. The table appears, it looks finished, and people leave before the link is written. The table is the middle of the answer. Wait for the link.

A good row

Every row is a company, and every column is separated by a pipe. This is a shipper row, which has ten of them:

Acme Precast | Edinburg, TX | Year-round citrus | Director of Logistics | Dana Reyes | (956) 555-0100 | acmeprecast.com | ops@acmeprecast.com | 1234567 | Opened a second yard in March

Nothing says which value is which except its position. That is what makes the three failures below possible, and it is also what lets us catch them.

One. A cell goes missing

The assistant has no named person, so instead of writing not found it writes nothing at all and moves on. Every value after it slides one column to the left:

Acme Precast | Edinburg, TX | Year-round citrus | Director of Logistics | (956) 555-0100 | acmeprecast.com | ops@acmeprecast.com | 1234567 | Opened a second yard in March

Read by position, the phone number is now the named person, the website is the phone number, and the USDOT is an email address. Every one of those is a plausible-looking value in the wrong place, which is the worst kind of wrong.

What we do: read the row by shape instead of by position. An email has an @, a website does not, a USDOT is five to eight digits, a phone is punctuation and digits. The values land in the right columns and the person is simply missing, which is the truth.

Two. The rows run together

Every column is separated by a pipe, and sometimes the assistant separates the rows with one too. Ten companies arrive as a single line of a hundred cells:

Acme Precast | Edinburg, TX | ... | Opened a second yard | Second Company | Salinas, CA | ... | Runs its own trucks | Third Company | ...

What we do: the arithmetic gives it away. A hundred cells where ten were expected, and a company name at every tenth position, is one answer rather than one company. We split it back apart and you get all ten.

Three. A pipe inside a name

Rarer, and the mirror image of the first one. A company whose own name contains the separator splits into two cells and pushes everything rightwards.

What we do: try rejoining the pair that makes the rest of the row make sense, and keep the version where the most columns hold the right shape of value.

What you will actually see

The import screen shows you the table before anything is saved, with a checkbox on every row. If something did not survive the trip it says so there, names the column, and offers to run it again rather than saving something quietly wrong. Your answers stay filled in, so a rerun is one press.

A rerun usually comes back clean. These are not failures of the search, they are formatting accidents on the way home, and asking the same question again generally produces a tidier answer.

If none of this matches

Send the link, or a screenshot of the answer, to colin@pettybuilds.com. Every one of the three above was found because somebody did that.