Stop Treating the Chat Box Like Google

Most managers prompt the way they search: a few keywords, hit enter, hope for the best. Then they complain the output is "generic." Of course it is. You gave it nothing to work with.

The job of a manager is to give context. That's literally what you do all day — translate goals into work, work into updates, updates into decisions. Prompting is the same skill in a smaller frame. If you can brief a new hire on a project, you can prompt a model. If you can't brief a new hire, fix that first.

Here's the mental model: every prompt has four parts. Role (who the model is acting as), context (what it needs to know), task (what you want it to do), and format (how the output should look). Skip any of these and you get sludge.

The Context-First Prompt

Power users obsess over clever instructions. Managers should obsess over context. The model already knows English. It does not know that your engineering team has six people, that two are on parental leave, that your VP cares about velocity but pretends to care about quality, or that your last reorg failed because of a Slack thread.

A prompt like "write me a status update" gets you a LinkedIn post. A prompt like the one below gets you something you can actually send.

You are my chief of staff. I run a 12-person product team
at a mid-size B2B SaaS company. My boss is the VP of Product;
she scans updates on her phone between meetings and wants the
headline first, risks second, asks third.

This week:
- Shipped the billing redesign (on time)
- Slipped the SSO integration by one sprint (vendor delay)
- Lost an engineer to a competing offer (backfill open)
- Customer NPS up 4 points

Write a Monday update. Under 150 words. No buzzwords.
Lead with the headline. Be honest about the slip but
don't catastrophize it. End with one specific ask.

Notice what's doing the work: the audience, the constraints, the tone, the length, and the structural rule (headline first). The bullets are facts. The model assembles. You edit.

If you find yourself rewriting the output heavily, the fix is almost always more context, not cleverer instructions.

Structured Outputs Beat Prose

Free-form prose is hard to act on. Force structure and the model gets sharper, and you get something you can paste into a doc, a ticket, or a deck without reformatting.

When you want a decision memo, ask for one explicitly:

Format as:
- Decision needed (one sentence)
- Options (3 max, each with one-line pros and cons)
- Recommendation (with the reason)
- What we'll know in 30 days that we don't know now

That last line is the secret. It forces the model — and you — to acknowledge what's being bet on. Most decision memos collapse under their own confidence. Yours won't.

For 1:1 prep, ask for a table:

Make a 4-column table:
| Topic | Why it matters to them | What I want to learn | My one open question |

For risks, force a severity column. For action items, force an owner and a date. Whatever the deliverable is, name the columns. The model will not invent the discipline you don't ask for.

Reusable Templates Beat Heroic One-Offs

Stop writing fresh prompts every Monday. Build three or four templates you use weekly and paste them with the variables swapped. This is the difference between people who say "AI saves me an hour a week" and people who actually mean it.

A status update template:

You are my writing partner. Audience: [WHO].
Tone: [TONE]. Length: [WORDS] words. Format: [STRUCTURE].

Facts to include (do not invent any others):
[BULLET FACTS]

Things to soften / things to escalate:
[NOTES]

A 1:1 prep template:

You are coaching me on a 1:1 with [NAME], my [ROLE].
Recent context: [3-5 BULLETS — projects, mood, last 1:1].
My goals for this conversation: [LIST].
Their likely concerns: [LIST OR "unknown"].

Generate:
1. Three opening questions in their own words
2. Two topics I'm avoiding that I shouldn't be
3. One thing to acknowledge before any feedback

A decision memo template:

Decision: [ONE-SENTENCE QUESTION].
Stakes: [WHAT CHANGES IF WE GET IT WRONG].
Constraints: [BUDGET, TIME, POLITICS].
What we know: [BULLETS].
What we don't: [BULLETS].

Produce: 3 options with tradeoffs, a recommendation,
and a kill-criterion (what would make us reverse this).

Keep these in a Notes file, a Notion page, or pinned in your chat tool. The investment pays back the first time you skip rewriting from scratch.

What Power Users Get Wrong

Power users tune their prompts like they're tuning a race car. They argue about chain-of-thought versus tree-of-thought. They paste in 2,000-word system prompts. For management work, most of that is wasted motion.

You are not building a product on top of the model. You are getting unblocked. The wins are:

• Brevity in, brevity out. Long rambling prompts produce long rambling answers. Edit your prompt down before you send it.
• One job per prompt. Don't ask for the update and the meeting agenda and the Slack message in one shot. The quality drops on all three.
• Show, don't describe. If you want a certain tone, paste an example of an update you liked. Models imitate examples better than they follow adjectives.
• Push back on the first draft. The second pass is almost always sharper. "Make it 30% shorter and remove anything that sounds like consulting" is a real instruction that works.

If you want a deeper dive into the broader manager-AI workflow, the AI for Managers playbook course covers the operating cadence end-to-end. For sharper written outputs specifically, AI writing and content creation is worth an evening.

The One-Page Discipline

Pick one task you do every week — Monday update, 1:1 prep, sprint review notes — and build the prompt template for it this week. Use it four Mondays in a row. Refine it twice. That's it. That single template will save you more time than any prompt-engineering course you'll ever pay for.

Prompting like a manager is not about being clever. It's about being clear, structured, and ruthless with context. The model is a fast junior who's read everything and remembers nothing about your company. Brief it accordingly.