I take notes like a person who fully intends to clean them up later and never does. A typical line from a conference looks like this: 'talked to Marcus Chen at Accenture, does ESG reporting, wants intro to our CEO, follow up re: the panel.' No structure, no consistency, sometimes a name with no company, sometimes a company with no name, occasionally a cryptic 'the tall guy from the fintech who hated our pricing.' Three months of that, across seven different docs in three different apps, roughly 200 lines. Then marketing asked me for a clean contact list for an event invite, and I realized my 'system' was a swamp wearing a trench coat.
Doing it by hand was a genuinely depressing prospect. Read a line, find the name, find the company, guess the role, copy it into a spreadsheet, repeat 200 times, lose my place around row 60, start a column over because I'd misnamed it, develop a headache, question my life choices. I estimated three hours of work and a small crisis of self-respect. I had also, at this point, read enough breathless posts about AI to be skeptical that any of it actually worked on real-world mess. So instead of the three hours, I tried the lazy thing, fully expecting to be disappointed and then do the three hours anyway.
One paste, one prompt, one table
I mashed all seven docs into a single block, pasted it into GPT-4o with the Messy Text to Clean Table, Zero Hallucinated Cells prompt, and asked for columns: Name, Company, Role, What We Discussed, Follow-up. GPT-4o is exactly the right tool here — fast, cheap, and very good at high-volume, low-glamour extraction where you just need it to be careful rather than creative. This is not a job that needs deep reasoning; it needs accurate, patient parsing at speed, which is GPT-4o's whole personality. It returned 47 rows in about 90 seconds, which is roughly the time it takes me to manually process two contacts and one existential sigh.
The part that turned this from a parlor trick into something I actually trust was the discipline around gaps. Where my notes never recorded someone's role, the cell said [NOT FOUND] instead of a confident guess. Where I'd written something ambiguous — was 'Priya, procurement, Deloitte' a Deloitte employee or someone who merely knows the Deloitte procurement lead? — it gave its best read with a (?) appended so I'd know to check. Two of those (?) flags caught misreads that would have put the wrong company on a follow-up email, the kind of small error that quietly tells a prospect you don't actually remember them.
- [NOT FOUND] cells became my to-do list: 12 contacts I needed to enrich before importing
- The (?) markers flagged exactly the two rows worth double-checking, not all 47
- EXTRACTION NOTES told me how many rows were incomplete — an honest report card on my note-taking habits
- CSV output dropped straight into the CRM importer with zero reformatting
- It added an 'Intro Requested' column on its own initiative and flagged why — three people had asked for the same intro and I'd never connected them
Why a marked gap beats a smooth lie
The single rule I'd tattoo on this prompt is 'a marked gap is correct, an invented value is a bug.' Earlier extraction attempts I'd tried with looser, friendlier instructions would helpfully 'complete' a row by inferring someone's title from their company — which is exactly how you end up confidently emailing a 'VP of Engineering' who is actually a six-week contractor, or addressing the intern as a director. I would much rather stare at a blank I can fill in than ship a confident fiction I won't notice until it's already embarrassing me in someone's inbox. GPT-4o follows that no-invention instruction tightly when you make it explicit and unambiguous, which is the entire point of writing the rule into the prompt instead of hoping for the best.
The extraction took 90 seconds. I then spent about an hour verifying the (?) rows and chasing down the [NOT FOUND] gaps — which is the right kind of work, the part that genuinely needs a human who was in the room, and it's a vastly better use of an hour than transcription. The math that matters: three hours of soul-crushing copy-paste became 90 seconds plus one hour of judgment. I pair this with my Plain-Language Paper Summary with Honest Caveats prompt when the 'messy text' is a dense report and I want both the gist and a structured pull of its actual numbers in one pass.
If your data lives in paragraphs, emails, Slack threads, or the chaotic margins of your own brain, grab the Messy Text to Clean Table, Zero Hallucinated Cells prompt on Prompt Dock, paste your mess in, and let GPT-4o hand you a table that tells you the truth — including the slightly humbling truth about everything you forgot to write down.