Working with LLMs from Zero
by Sergey Okinchuk

Chapter Five · Output That Goes Somewhere

Getting Usable Output

An answer is text you read; an artifact is a thing that slots into the place it was always headed. Four pages turn the first into the second: ask for the artifact, ask for a shape a machine can unpack, save the prompt that produces it every time, and check the result before anyone acts on it. By the end Tessa has one file holding 2,300 classified reviews — and a way of saying what she checked.

Chapter 4 got Waymark's material in front of the model, and what came back was prose. Good prose, correct prose — and prose that Tessa then spent twenty minutes a time retyping into a tracker, a template, an email. That retyping is the subject of this chapter, and it turns out to be avoidable by saying one more sentence in the prompt.

The four pages build one workflow, in order. Ask for the artifact rather than the answer, by naming where the output is going. Ask for it in a shape a program can unpack without guessing — which is JSON, and it takes ten minutes to learn to read. Save the prompt that produces that shape, so consistency stops depending on what anyone remembers on a Thursday. Then check what comes out, because every field in every cell is still generated text, and the hotel that does not exist is about to reappear inside a column. Underneath all four is a quiet second story: work with a fixed prompt, a slotted input and a collected output has the exact shape of a program — which is what Chapter 8 will make of it.

Two things that can come back from the same question
An answer to read
A clear paragraph about last week's complaints — correct, well organized, and still a long way from the tracker.
Output to use
The same knowledge as labelled rows in the columns the tracker expects, checked against a sample, pasted onward untouched.

Topics in This Chapter

Topic 21
Output You Can Reuse
Six seconds of thinking, twenty minutes of reshaping — the bottleneck nobody names. Answer versus artifact, why naming the destination beats describing the format, and what identical shape across 2,300 results actually buys you.
Concept
Topic 22
Text a Program Can Read
The tracker imports structured text, and the word for it is JSON. Labelled boxes rather than programming: braces, labels, values, lists — read in ten minutes, and the same shape that turns up inside the request in Chapter 7.
Concept
Topic 23
Templates and Repeatable Prompts
Monday's prompt rewritten from memory on Thursday gives different columns and a file the tracker rejects. Fix the constant, slot the variable, version the document — and notice what shape the resulting afternoon has.
Concept
Topic 24
Check the Output, Don't Trust It
Two hundred rows in, a cell names the Harbourview Annex. Why structure launders errors, how a random dozen tells you about 2,300, which checks need no judgement at all, and how a found error becomes a better template.
Concept