Part 2 · Control
Output control
The concept
Most of the time you waste on AI output isn't spent getting a wrong answer — it's spent reshaping a right answer that came back in the wrong form. That's a solved problem: whatever shape you want, ask for it explicitly, and be numeric wherever you can. "Short" is an opinion; "maximum 60 words" is an instruction. The four dials you control are format, length, tone, and density, and each responds to precision. A useful mental model: the model will always produce something well-formed — you're choosing which well-formed thing, and if you don't choose, it chooses for you (and it defaults to a chatty, hedge-heavy, bullet-happy house style).
Format. Say the container. Markdown with named headings. A table with these exact columns. A numbered list of exactly 8. JSON with these keys. A single paragraph with no line breaks. Slide-by-slide with a headline and three bullets each. If you're going to paste it into PowerPoint or a spreadsheet, ask for the format that pastes cleanly — tables for spreadsheets, Slide N: headline / bullets blocks for decks.
Length. Always a number, and pick the right unit. Words for copy. Sentences for emails. Characters for ad headlines and captions. Slides for decks. Bullets for summaries. "One paragraph, 4 sentences maximum" beats "briefly." And know that models are systematically bad at exact word counts — they'll land within about ±15%. If exactness matters (a 30-character banner headline), ask for the count and ask it to print the count after each option so you can verify at a glance.
Tone. Three adjectives plus one anti-adjective. "Confident, warm, plain — not salesy." The anti-adjective does most of the work, because it kills the specific failure mode you're worried about. Better still: name a reference. "Sound like an Economist leader, not like LinkedIn."
Density. This is the underused dial. "No filler, no preamble, no summary of what you're about to do, start with the answer" removes about a third of the words from most responses. Add it to everything you write for your own consumption.
Two commands worth memorising:
No preamble. Start with the deliverable.— kills "Certainly! Here's a great..."If you're unsure about a fact, mark it [VERIFY] rather than guessing.— see Module 12.
Before → After
Summarise this client meeting transcript.What you get: eight paragraphs of narrative, decisions buried in the middle, action items implied but not extracted, and you re-read the whole thing anyway.
Summarise this client meeting transcript.
No preamble. Output exactly these four sections:
## Decisions Made
Bullets. Each one line. Only things actually agreed — not discussed.
## Open Questions
Bullets. Each with who needs to answer it.
## Action Items
A markdown table: | Action | Owner | Due | Blocked by? |
Owner must be a named person from the transcript, or "UNASSIGNED".
## Risks I Should Worry About
Max 3 bullets, your judgement, one line each. If none, write "None."
If something was discussed but not resolved, it belongs in Open
Questions, not Decisions. Do not infer decisions that weren't stated.A quick reference card
| You want | Write this |
|---|---|
| Something pasteable into Excel | "Output as a markdown table with columns X, Y, Z. No commentary." |
| Something pasteable into a deck | "Slide 1: [headline ≤7 words] / 3 bullets ≤12 words each. Repeat for 8 slides." |
| A tight email | "Under 120 words. Subject line + body. One idea per paragraph." |
| Ad headlines | "15 options, each ≤40 characters. Print the character count after each." |
| A decision, not options | "Give me your single recommendation first in one sentence, then the 2 runners-up and why you rejected them." |
| Data your system can read | "Return only valid JSON, no markdown fences, keys: ..." |
Do this
Take a task you do weekly — a meeting summary, a status report, a WhatsApp update to a client. Write the output-format block for it once, properly. Save it. You have just built the first entry in your prompt library (Module 13), and you will never format that deliverable by hand again.