The Kickoff Document That Cost Me Two Extra Hours

One prompt. Two jobs. Only one of them worked, and the broken half was invisible.

I gave one AI platform a single instruction: research the market and write the kickoff document. It did both. The output looked finished.

It took me two extra hours to find out it wasn't.

The instruction had two jobs in it

The prompt asked for competitive research and a written deliverable in the same breath. That reads as one task to a person, because a person doing it would research first, form a view, then write. The model doesn't sequence that way. It produced a document shaped like the thing I asked for, and the research half arrived already folded into the prose.

Which meant there was no point at which I could inspect the research on its own. It came pre-argued.

Real content from the wrong market

The failure wasn't hallucination. Every fact in the research section was true. It just belonged to an adjacent market — similar product category, different buyer, different constraints.

That's the part that cost the two hours. Invented content fails a plausibility test almost immediately; something is off, the numbers don't sit right, a claim doesn't survive a search. Real content from the wrong market fails no plausibility test at all. It reads correctly because it is correct, somewhere else.

I only caught it when a figure in the document contradicted something I already knew about our own buyers. Then I had to re-verify the whole section, line by line, because once one item is from the wrong place you can't assume the rest aren't.

Composite prompts hide their own failures

A single-job prompt produces output you can check against the job. Ask for research, and you get a list you can verify item by item. Ask for a document, and you can read it against the brief.

Ask for both, and the model resolves the ambiguity for you. It decides how much research is enough, decides which findings matter, and then buries those decisions inside finished prose. You don't see the judgment calls because they've already been written past.

The output looking finished is what makes it dangerous. A rough list invites scrutiny. A polished document invites signature.

Split the task, not the tool

The fix isn't a better model or a longer prompt. It's two passes.

Pass one — research only. Ask for findings as a table, with a source and a market label on every row. Explicitly instruct it not to write prose and not to draw conclusions. You want the raw material in a form where a wrong row is visible as a wrong row.

Pass two — the document. Paste back only the rows that survived your review, and ask for the deliverable built from those and nothing else.

The second pass takes about four minutes. The first takes ten, most of it spent reading. That's still faster than one composite prompt plus two hours of forensic re-verification, and it's dramatically faster than finding the error after the document has been circulated.

What to watch for

Any prompt containing the word "and" between two verbs is a candidate. Research and write. Summarize and recommend. Analyze and prioritize. Each pair looks like one task and behaves like two, and the seam between them is where the unverifiable work hides.

The kickoff document was fine in the end. It just cost three hours instead of one, and the extra two bought nothing except the knowledge that I'd asked one question when I should have asked two.

A version of this piece first appeared in AI in Plain English on Medium.