Give context, constraints, and examples

Chris typed “write a short update for leadership” into Claude at 8:41. The huddle on the calendar was labeled “ops standup (90s)” and started at 9:00. The reply came back at 1,140 words, with H2s named Strategic outlook, Cross-functional alignment, and Recommended next steps for FY27. Chris needed about 120 spoken words on last week’s 14 closed tickets, 3 still open, and the Tuesday login outage that ran 37 minutes from 14:12 to 14:49 PT. The VP was already in the Zoom waiting room. The sticky on the monitor still said keep it to tickets.

This is Part 13 of Phase P, and Part 2 of Prompting for everyone. Part 1 was talk like a clear human: a real task in ordinary words, no magic phrases. A clear task is still not a brief. Models fill the hole with a default memo, a default reader, and a default future. This part packs the missing pieces: context (who, why, facts), constraints (length, tone, must-nots), and examples (one good snippet of the shape). Next is iterate when the first answer is wrong. Accounts and billing live in AI setup from zero (Free vs paid). The door chooser is Which AI product should I use?. Everyday writing jobs sit in AI for writing and questions.

What you’ll learn

• Why “short update for leadership” still produces a three-page vision memo
• How to write context as who, why, and the facts the model cannot invent
• How to set constraints as a length cap, a tone, and a must-not list
• Why one tiny example of the output shape beats a pile of adjectives
• A packed prompt you can copy for Chris’s 90-second standup, then reuse on a different reader with the same facts

The model fills the hole you leave

ChatGPT, Claude, Gemini, and Grok are good at completing a pattern. They are less good at knowing which meeting you are walking into. “Short” could mean 80 words or 800. “Leadership” could mean the VP who already sat through the incident channel, or a board packet that wants a thesis. “Update” is a noun with no slot. The model still has to emit something. So it reaches for the shape it has seen a thousand times: titled sections, a horizon, three recommendations, a warm close.

Chris did not ask for FY27. The hole asked for it. There was no audience, no clock, no ticket counts, and no ban on strategy language. A 90-second standup has a clock. A vision memo does not. If you do not name the clock, you get the memo.

You will see the same fill on other jobs. “Help me study this chapter” becomes a 12-heading outline with no quiz. “Draft a note to the vendor” becomes a two-page relationship letter. The task was typed. The brief was empty. As of writing, vendor docs say this in their own labels. OpenAI’s ChatGPT Learn page talks about goal, context, output, and boundaries. Google’s Gemini tips talk about persona, task, context, and format. Anthropic’s prompting notes push you to add motivation and examples. None of them ask you to collect secret words. They ask you to stop making the model guess the meeting.

Rule of thumb: If you cannot point at who will hear it, why it exists this morning, and one fact the model cannot look up, you are still sending a task, not a brief.

Context: who, why, and the facts

Context is the part of the brief that lives in your head and nowhere in the prompt. Who is in the room. What they already know. Why this artifact exists now. Which numbers are allowed to appear. The model cannot see the sticky, the Zoom waiting room, or the Slack huddle title. If you skip those, it invents a generic executive and a generic quarter.

For Chris, useful context was small and local:

• Audience: VP Ops plus two directors. They already know the outage happened. They do not need a recap of what login means.
• Why now: the 9:00 huddle is 90 seconds. The goal is status, not a new program.
• Facts: 14 tickets closed last week, 3 still open (one waiting on a vendor patch), Tuesday login outage 37 minutes, no customer data lost, status green as of 17:10 PT Tuesday.

That is enough. You do not paste the whole incident doc. You paste the facts that would change the spoken sentence. “No customer data lost” changes the tone. “Vendor patch” changes the ask (there is no ask). “They already know” stops the model from opening with “Yesterday, our systems experienced…” as if the room were cold. If you omit the 37 minutes, some chats will fill with a round number that sounds official. Put the facts in the prompt so the draft cannot wander into a press release.

Do not dump the last six months. Context is the slice that changes this output. Chris’s FY27 hiring plan is real. It does not belong in a 90-second standup prompt unless the huddle is about hiring. Extra context is how you invite the vision memo back in. If a fact would be weird to say out loud in that room, it probably should not be in the prompt for that room.

Constraints: length, tone, and must-nots

Constraints are the fences. Length is the one you can check with a timer. Tone is the one you can check by reading the draft aloud. Must-nots are the ones that would create extra work if they slipped through: a fake deadline, a new OKR, a promise you cannot keep in Slack.

Adjectives are weak fences. “Keep it short.” “Be professional.” “Make it executive-ready.” Those words are compatible with 1,140 words and three H2s. A checkable fence looks like this:

• Length: 90 to 120 words, meant to be spoken in 90 seconds. No headings. No bullet list in the spoken version.
• Tone: calm, operational, present tense where it is true. No “exciting opportunity,” no “strategic,” no “alignment.” Chris talks like a support lead, not a keynote.
• Must-not: no FY27 roadmap, no hiring ask, no new OKRs, no customer names, no invented numbers. If a number is missing, say “I do not have that” instead of guessing. If your org mocks “circle back,” put that on the list too.

OpenAI’s consumer prompting page (as of writing) calls a close cousin of this “boundaries”: keep approved dates unchanged, flag missing information, prepare a draft and do not send. Google’s Gemini API notes tell you to put limits on length and to say what the model should not do. You do not need their vocabulary. You need a fence the draft can fail. 1,140 words fails 90 to 120. “Recommended next steps for FY27” fails the must-not list.

Put the must-nots in a short list, not buried in a paragraph of vibes. If the output still smuggles a roadmap, you will fix that in Part 3 by pointing at the line. The first prompt should still name the ban. Iteration is cheaper when the fence was there. “Punchy” and “board-ready” are guesses. “Spoken, no headings, you can hear it in 90 seconds” is a test.

Examples: one good snippet beats adjectives

An example is a tiny sample of the shape you want, not a second copy of the facts. Anthropic’s prompting notes (as of writing) treat examples as one of the more reliable ways to steer format and tone. Google’s Gemini API docs recommend few-shot examples in the prompt. You do not need five. One good spoken paragraph from last week is enough for a standup. Two is plenty if the second one shows a miss: a version that is too long, or that sneaks in a hiring line.

Chris had a real sample from the week before, 94 words, no headings, one number cluster, zero asks. That sample did more than “make it concise.” Concise is a feeling. 94 words is a ruler. Paste the sample and label it as shape, not as this week’s facts, so the model does not copy last week’s ticket counts into this week’s mouth. If your sample ends with “Ask: none,” you will hear fewer invented asks.

Skip stock “act as a world-class chief of staff” wrappers. They are adjectives in a costume. If you do not have last week’s script, write three sentences by hand that you would say out loud, and paste those. Ugly and true beats polished and generic.

Pack all three into one prompt

Here is Chris’s hole next to the packed brief. Same task verb. Different Monday.

Copy this into ChatGPT, Claude, Gemini, or Grok. Swap the facts. Keep the labels. As of writing you do not need XML tags for a huddle script; labeled sections in plain English are enough for consumer chat. (API docs may show XML or fences. Use those when you are wiring an app. Use this pack when you are late for Zoom.)

TASK
Draft the 90-second standup script I will say at 9:00.

CONTEXT
Audience: VP Ops plus two directors. They already know Tuesday's login outage happened.
Why now: weekly ops huddle, 90 seconds, status only.
Facts you may use (do not add others):
- 14 tickets closed last week
- 3 still open; one is waiting on a vendor patch
- Tuesday login outage 14:12 to 14:49 PT (37 minutes)
- No customer data lost
- Status green as of 17:10 PT Tuesday

CONSTRAINTS
Length: 90 to 120 words. No headings. No bullet list in the spoken script.
Tone: calm, operational, support-lead voice. No "strategic," "alignment," or "opportunity."
Must-not: no FY27 roadmap, no hiring ask, no new OKRs, no customer names, no invented numbers.
If a number is missing, write "I do not have that" instead of guessing.

EXAMPLE OUTPUT (shape only; last week's facts, not this week's)
"Last week we closed 11 tickets and carried 2. The only incident was a 12-minute cache miss on Thursday. Status: green. Ask: none."

What that prompt is aiming at, this week’s facts, spoken:

Last week we closed 14 tickets and still have 3 open. One open item is waiting on a vendor patch. The only incident was a 37-minute login outage Tuesday from 14:12 to 14:49 PT. No customer data lost. Status green as of 17:10. Ask: none.

Read that second block aloud. It is about 70 words. Under 90 seconds even if Chris adds one breath. That is the product. The 1,140-word memo was a different product wearing the same task verb.

Same facts, different pack

The three pieces move when the reader moves. Keep the outage numbers. Change the slot to a customer-facing status note. Context becomes “public post, customers already saw the 37-minute blip, legal already approved ‘no data lost.’” Constraints become “under 80 words, no internal ticket counts, no vendor name, no humor.” The example becomes last quarter’s status blurb, not last week’s huddle script. Paste a new pack. Do not reuse the standup script and hope the model knows it is public now. It does not know. You have to say who is outside the Zoom.

Common mistakes

• Stopping at a clear verb. “Write a short update” is a task. The 1,140-word memo is what a missing brief looks like.
• Using adjectives as fences. “Short,” “punchy,” and “executive” all fit a three-page vision note. Use a word cap or a timer.
• Pasting six months of Slack as “context.” Extra history invites extra recommendations. Slice the facts that change this sentence.
• Asking for “an example” and then describing the vibe. Paste a snippet of the shape, even a rough one you wrote by hand.
• Letting the example’s old numbers leak into this week. Label the sample as shape only, then list this week’s facts in their own block.
• Skipping must-nots because you assume the model “would never” add a hiring ask. Chris’s FY27 section arrived anyway.

How to practice this week

Pick one real artifact you will send in the next two days: a standup script, a status note, a parent email, a vendor ping. Write the thin one-liner you would have typed. Then fill the four labels from the code block (task, context, constraints, example). Run both in the same chat tool you already use. Time the spoken draft. If it blows the cap, tighten the constraints before you rewrite the whole thing by hand. Keep the thin prompt in the thread so you can see the hole. Next in this series is when the first answer is still wrong. You will need that once the brief is packed and a line still drifts. Paths back: Learn, Practical AI, and Prompting for everyone.

Quick recap

• A clear task still leaves a hole. The model fills it with a default memo.
• Context is who, why, and the facts the model cannot invent.
• Constraints are a checkable length, a speakable tone, and a must-not list.
• One real snippet of the shape beats “make it punchy.”
• Pack all three around the verb before you hit send. Change the pack when the reader changes.

Sources

• OpenAI, ChatGPT Learn: Prompting (goal, context, output, boundaries; as of writing)
• OpenAI: Prompt engineering (API-side prompting notes; names move)
• Anthropic: Prompt engineering overview (when prompting is the right lever)
• Anthropic: Prompting best practices (context, examples, structure)
• Google: Gemini API prompt design strategies (few-shot examples and constraints)
• Google Workspace: Tips to write prompts for Gemini (context, natural language, split hard jobs)
• Anthropic: Getting started with Claude
• AMS: Prompting for everyone (this series)
• AMS: Talk to AI like a clear human (Part 1)

Merch on Etsy

Open store →

Y Hat Mouse Pad, Data Science Mouse Pad, Analytics Mouse Pad, Y Pred Mouse Pad View on Etsy → Good Cheap Fast White Ceramic Coffee Mug, Data Science Mug, Data Mug, Analytics Mug, Statistics Mug, Programming Mug, Gift Mug, Coffee Mug View on Etsy → Imbalanced Classes Mouse Pad, Data Science Mouse Pad, Analytics Mouse Pad, Imbalanced Mouse Pad View on Etsy → Naive Bayes Mouse Pad, Data Science Mouse Pad, Analytics Mouse Pad, Bayesian Mouse Pad View on Etsy →