Templates and Repeatable Prompts
On Monday Tessa runs fifty reviews through the chat box and the result is the best hour of her week: labelled fields, four clean buckets, straight into the tracker. On Thursday she sits down to do the next fifty, types the request again from memory, and gets something else. The category names have drifted. There is an extra column she never asked for. The ratings are written as words rather than numbers, and the tracker rejects the file.
Nothing broke. Thursday-Tessa simply wrote a different prompt from Monday-Tessa, because she was working from memory and memory does not store punctuation. The fix is the least glamorous idea in this book and one of the most valuable: write the prompt down once, keep it, and change only the part that is genuinely new each time.
Chapter 2 already gave that its name: a saved prompt with a gap in it for the input is a template. What that chapter planted as a habit, this one turns into a working document. Think of the office's letterhead. Nobody re-invents the covering letter each morning — the layout, the boilerplate paragraphs and the sign-off are fixed, and you change the recipient and the one paragraph that is actually new. That is exactly the move, applied to a chat box.
Fix the Constant, Slot the Variable
Building a template is one decision made carefully: what stays the same every single time, and what changes.
What stays the same is nearly everything, and it is everything Chapter 2 taught. The instruction — what to do with each review. The bucket definitions, spelled out so they cannot drift. The list of fields and their labels, from the last page. One worked example, filled in by hand, showing exactly what a finished item looks like. The rule for what to do when a field has no honest answer. All of that is fixed text, written once.
What changes is one thing: the reviews. So the template ends with a marked-off gap — a line saying the reviews follow, and then the reviews. Same instruction in, same shape out, run after run.
Same shape, to be precise about it. Chapter 1 established that the model does not repeat itself word for word, and it does not: two runs of the identical template over identical text can word a summary differently. What holds is the structure — the fields, the labels, the four buckets — because the structure was shown rather than left to chance. Pinned shape with varying wording inside it is not a compromise, it is the arrangement you want.
The Template Is a Document
Once the prompt is written down, treat it like any other working document at Waymark, because that is what it now is.
It lives where the team keeps documents, not in one person's chat history. It has a name that says what it does. And it has a version, because it will change: when Tessa adds a field for whether a review mentions a refund, the note at the top reads v3: added the refund-mentioned field. Six weeks later, when a file of results looks odd, that line is what tells her which template produced it.
The sharing is the part people skip and the part that pays. When the template is a document, a colleague covering for Tessa produces a Thursday batch that merges cleanly with her Monday one — same fields, same buckets, same rules for a vague review. Without it, Waymark has two people improvising two similar prompts and quietly generating two incompatible files.
This is what turns the craft of Chapter 2 from a personal skill into something the office owns. A prompt that only works when a particular person remembers it is not a process; it is a habit that leaves when they do.
Running Batches Through It
Now the promise of this chapter's first page gets delivered. Fifty reviews into the gap, run, collect the output. Fifty more, run, collect. Every run shares the shape, so the outputs stack into one file without anyone reconciling anything, and by the end of the afternoon the 2,300 reviews that defeated Waymark for years are one file with five labelled fields per row.
Keep the batches small, for the reason Chapter 2 gave: over a long output the shape drifts near the end, columns merging and labels wandering. Fifty at a time with the full template re-shown each run beats two hundred in one go, and takes less total time than repairing the drift by hand. What was a tip in Chapter 2 is now policy, because a drifted batch does not just look untidy — it fails to import.
One Thing Worth Noticing
Look at the afternoon Tessa just had, and describe the work rather than the result.
A fixed piece of text. A gap that gets filled with the next fifty items. A run. An output collected into a growing file. Then the same four steps again, and again, forty-six times, with a human being supplying nothing but the copying and pasting.
Work with that shape — fixed instruction, slotted input, collected output, repeated — is not really work for a person. It is the exact description of what a program does, and a very short program at that. Chapter 8 opens with Tessa's contractor writing it, in twenty lines she can read. Everything that program will do, she has already done by hand this afternoon; the template is its blueprint.
Before that, one thing stands between a stacked file and a file anyone should act on. Every field in those 2,300 rows was generated, and generated text has a habit this book named in Chapter 3 and has not forgotten.
- "I will remember what I prompted." Thursday-Tessa is the counterexample, and she had only three days to forget in. Unrecorded prompts drift, and drifted shape breaks everything downstream of it.
- "A template kills the model's flexibility." It pins the shape, not the thinking. The summaries still respond to each review; they just arrive in the same columns, which is precisely the combination you want.
- "Templates are an engineering practice." They are an office practice with a long history. This is the boilerplate letter and the pre-printed form, pointed at a chat box.
- "A saved template will keep working unchanged forever." It is a document, so it gets edited — a new field, a tightened bucket definition. That is why it carries a version note, and why the note says what changed.
- Templates turn prompting from a personal talent into a team asset — the first step of any workflow that has to outlive one person's memory of what they typed.
- The template is the automation blueprint. What Chapter 8 hands to a program is this document with the copying and pasting taken over, which is why the shape of this afternoon is worth noticing.
Knowledge Check
In a template, which part is the "slot"?
- The four bucket definitions, reworded each run to suit the batch in hand
- The worked example, swapped for a fresh one drawn from each new batch
- The reviews themselves, the one thing that genuinely differs between runs
- The list of field labels, adjusted to whatever each batch of reviews happens to mention
Why does the page insist a template be a shared, versioned document?
- So the model can see the history of edits and take the improvements into account next run
- So the team can show who wrote which prompt if the output is ever questioned
- So the prompt stays short enough to leave room in the context window for the reviews themselves
- So a colleague's run matches Tessa's, and an odd file can be traced to the version that made it
Why run fifty reviews at a time rather than two hundred?
- Over a long output the shape drifts near the end, and drifted output will not import
- Larger batches cost noticeably more to run each time, so smaller ones keep the bill down
- The model gets less careful as a single request grows longer and more repetitive
- Fifty rows is the largest number a person can reliably check by reading through them all
What does the page say Tessa's afternoon of batches reveals about the work?
- That classifying reviews is judgement work, so a person has to stay in the middle of it
- That its shape, fixed text with slotted input and collected output, is a program
- That the chat box cannot handle work at this volume and a different tool is needed
- That templates lose their usefulness once the number of batches gets large
You got correct