3. Controlling the Shape of What Comes Out
We asked for one set of numbers in three formats. The table came back with four rows when we only supplied three facts, and the fourth was a metric it derived itself.
The WJS Desk
Sep 7, 2026 · 7 min read
Most people accept whatever shape the answer arrives in and then reformat it by hand. That is a strange habit, because format is the single easiest thing to control and the model has no idea what you plan to do with the output.
A status update going into Slack, a table going into a deck, and three bullets going on a board slide are the same facts in three shapes. You can ask for all three at once.
One set of facts, three shapes
We gave it a genuinely boring set of numbers, the kind you actually have to report on:
Our team ships a release every two weeks. Last quarter we
shipped 6 releases, 2 slipped by a week, and we had 3
rollbacks.
Give me this three ways, labelled:
1. As a single sentence for a Slack channel.
2. As a 4 row table with a Metric and Value column.
3. As three bullet points for a board slide, no more than
8 words each.What came back, in full:
Slack: Last quarter we shipped 6 bi-weekly releases, though 2 slipped by a week and we had 3 rollbacks that impacted our velocity.
| Metric | Value |
|---|---|
| Releases Shipped | 6 |
| On-Time Delivery | 67% (4 of 6) |
| Releases Delayed | 2 (1-week slip) |
| Rollbacks | 3 |
Board bullets: Delivered 6 bi-weekly releases successfully · 2 slipped by one week · 3 rollbacks affecting overall stability
Three usable outputs, one prompt, about twenty seconds. But look at the highlighted row, because that is the lesson inside the lesson.
The format made it invent a number
We gave it three facts: 6 releases, 2 slipped, 3 rollbacks. The table has four rows. The extra one is On-Time Delivery, 67% (4 of 6), and we never supplied it.
The arithmetic is right. Four of six is 66.7%. It is a reasonable derivation and in this case a useful one. It is also a number that did not exist until we asked for a four row table, and if you paste that table into a deck you are now presenting a metric you did not calculate and did not check.
Asking for a shape creates pressure to fill it. We asked for four rows and gave it material for three. It found a fourth. That is exactly the behavior you want when you are brainstorming and exactly the behavior that puts a wrong number in front of your VP. If you specify a row count, a bullet count or a word count, check whether the output was padded to reach it.
The fix is one clause: "use only the figures I gave you, do not derive new metrics". Or ask for "up to 4 rows" rather than "4 rows", which removes the pressure entirely.
The four dials
Length. The most under-used instruction there is. "Under 150 words." "One paragraph." "No more than 8 words each." Without a limit it writes to a comfortable middle length that is almost always too long for wherever you are pasting it.
Structure. Name the container: a table with these columns, a numbered list, three bullets, a single paragraph, a checklist I can print. If you know the shape of the hole, describe the hole.
Register. Where is it going? "For a Slack channel" and "for a board slide" produced visibly different sentences above from identical facts. Naming the destination does more than naming a tone, because a destination implies an audience, a length and a formality all at once.
What to leave out. The dial nobody touches. "No preamble." "Do not restate the question." "No closing summary." "Do not explain your reasoning unless I ask." Every one of those removes a paragraph you were going to delete anyway.
The instruction that saves the most time
If you only take one thing from this lesson:
Give me just the output. No preamble, no explanation,
no closing summary.The default behavior is to wrap the thing you asked for in an introduction and a conclusion, because that reads as helpful. When you are going to paste the result somewhere, both are pure friction. This single line turns a response you have to edit into one you can copy.
Asking for several formats at once
This is the trick from the test above and it generalizes further than people expect. When you do not know which shape you want, ask for the shapes and pick:
Give me three versions, labelled:
1. One sentence, for someone who only reads the subject line.
2. One paragraph, for the team.
3. Full detail, for the person who will actually do the work.Three versions cost the same one request. You will usually take the middle one and steal a phrase from the first.
Getting it into other software
Say the destination format explicitly and it will comply:
- Spreadsheet: "as CSV, with a header row, comma separated, no other text"
- A document: "as markdown with H2 headings"
- A form or a tracker: "as a numbered list where each line is Field: Value"
- Something you will paste into a chat: "plain text, no markdown, no asterisks"
That last one matters more than it sounds. Markdown asterisks pasted into a plain text field become literal asterisks, and half the "why does the output look broken" complaints are that.
Say where it is going, not just what it should look like
Compare two ways of asking for the same constraint:
| Describing the format | Naming the destination |
|---|---|
| Keep it short and casual | This is going in a Slack channel |
| Formal, bullet points, high level | This is going on a board slide |
| Detailed, technical, thorough | This is going in a runbook someone reads at 3am |
| Friendly but professional | This is the first email a new customer gets from us |
The right column is shorter and does more, because a destination carries an audience, a length and a register together. "Someone reads this at 3am" implies short sentences, no preamble, and the command they need first, and you did not have to say any of that.
In our three format test above, "for a Slack channel" and "for a board slide" produced visibly different sentences from identical facts. The Slack version used "though" and ran the caveat into the same sentence. The board version broke into three parallel fragments under eight words each. We specified the word limit on the bullets, but not the parallel structure. The destination did that.
Iterating on shape without regenerating
The mistake that costs the most time is treating a badly shaped answer as a failed request and starting over. Format is the cheapest thing to fix in place, because the content is already there and only the container is wrong.
Same content, half the length.
Same content, as a checklist.
Same content, but for someone who has not read the thread.
Turn the third paragraph into a table.
Cut everything that is not a decision or an action.Five words each. Every one of them keeps what it already worked out and changes only the presentation. Rerolling the original prompt throws that away and gives you a different average.
The one worth memorising is the last: cut everything that is not a decision or an action. Applied to meeting notes, a status update or a long thread summary, it usually removes 60% of the text and loses nothing you needed.
Two shapes people never ask for and should
The diff. When you have edited something and want to know what actually changed in substance:
List what changed between these two versions, as a table
with What Changed and Why It Matters. Ignore rewording.The inverse. When you want to know what is not there:
What is missing from this that a reader would expect,
as a list, most important first. Do not comment on what
is already present.That second one is genuinely hard to get any other way. Errors of omission are invisible when you proofread, because nothing on the page is wrong. Naming the shape you want, a list of absences, is what makes them appear.
Common mistakes
- Specifying an exact count you cannot fill. As above. Use "up to N" unless you are certain the material supports N.
- Asking for a format without saying what goes in the columns. "As a table" leaves the columns to it. "A table with Metric and Value" does not.
- Reformatting by hand instead of asking again. "Same content, as a checklist" is five words and beats ten minutes of editing.
- Forgetting that format persists. Once it starts producing tables in a conversation, it tends to keep going. Say "prose from here" when you want to stop.
Try it now
Take the last thing you asked AI for and ask for it again in the shape you actually needed, plus "no preamble". Then look for anything in the output that was not in your input. On our four row table, one number in four was derived rather than given.
Next
Lesson 4 is the most famous prompting trick of all, telling it to think step by step. We tested it on two problems and got a result that surprised us.
