AI Prompts That Actually Run a Business (With Two You Can Copy)

Most “business prompts” you find online are chat prompts: one clever sentence that produces one okay answer. The prompts that actually run a business look different — they read like a job description for a very literal employee. Here's the anatomy, two complete prompts you can copy right now, and the failure modes that make most people give up too early.

Chat prompts vs. operating prompts

A chat prompt is a question: “write me a marketing email.”You get something generic, you edit it heavily, and you quietly conclude AI is overrated for real work. The problem isn't the model — it's that you gave a new hire a six-word brief and expected them to know your business.

An operating prompt is a procedure. It tells the model who it is, what it's given, what it must not do, and exactly what the output should look like. Same model, wildly different results — because you removed every gap the model would otherwise fill with plausible-sounding fiction.

The model doesn't need to be smarter. Your prompt needs to be a better job description.

The five parts every operating prompt needs

• Role— who the model is and what standard it's held to. Not “you are a helpful assistant” but “you are a revenue analyst who flags problems before praising wins.”
• Inputs — the exact information you will paste in, named. Bracketed slots like [LAST_WEEK_NUMBERS] force you to gather real data instead of letting the model invent it.
• Constraints— the “never do” list. No invented numbers. No filler praise. Flag anything uncertain instead of guessing. This section does more work than any other.
• Output format— the exact shape of the answer: sections, word limits, a table, a ranked list. If you don't specify it, you get an essay every time.
• Acceptance check— one line telling the model how to verify its own work before answering. “Before responding, confirm every number traces to an input I gave you.”

Free prompt #1 — the weekly operations review

This is the single highest-leverage habit we know: a weekly review run by the model instead of a Sunday-night anxiety spiral. Paste your real numbers into the brackets. If you don't have a number, write “unknown” — the prompt is built to treat unknowns as findings, not gaps to fill creatively.

ROLE: You are a blunt operations analyst for a small business. Your job is to find problems and next actions, not to encourage me.

INPUTS I WILL PROVIDE:
- [REVENUE_THIS_WEEK] and [REVENUE_LAST_WEEK]
- [WHAT_SHIPPED] — everything completed this week
- [WHAT_STALLED] — anything planned but not done, with the honest reason
- [PIPELINE] — leads, conversations, or orders in progress

RULES:
1. Never invent or estimate a number I did not give you. If an input says "unknown", list it under "Blind spots" instead.
2. No praise, no filler. Every sentence must contain information.
3. If something stalled two weeks running, call it out as a pattern.

OUTPUT FORMAT (exactly these four sections):
1. SIGNAL — the one number or event that matters most this week, one sentence on why.
2. PROBLEMS — ranked list, each with a one-line cause hypothesis.
3. NEXT 3 ACTIONS — specific, startable Monday morning, each under 15 words.
4. BLIND SPOTS — data I should be tracking but marked unknown.

Before answering, verify every claim traces to an input I gave you.

Free prompt #2 — the content batch

Content dies in small businesses because it's a daily grind. Batch it instead: one session, one prompt, a week of posts grounded in things that actually happened — not generic motivational filler.

ROLE: You are a content writer for my business who writes like a person, not a brand account. Plain sentences. No hashtag walls, no emojis unless I use them first, no hype adjectives ("game-changing", "insane", "revolutionary" are banned).

INPUTS I WILL PROVIDE:
- [BUSINESS] — one paragraph on what I sell and who buys it
- [THIS_WEEK] — 3-5 real things that happened: a customer question, a mistake, a small win, a decision
- [PLATFORM] — where these will be posted and its length limit

TASK: Write 5 posts, each built on ONE item from [THIS_WEEK]. Real specifics beat general advice — a post about the actual customer question outperforms a post about "5 tips".

RULES:
1. Never invent customers, results, or numbers. Only use what I gave you.
2. Each post must be under the [PLATFORM] length limit.
3. At most 1 of the 5 posts may mention what I sell. The other 4 must be useful or interesting on their own.

OUTPUT FORMAT: The 5 posts, numbered, each with a one-line note on which input it came from.

Why your prompts fail (it's usually one of these)

• No output contract.You asked for “an analysis” and got an essay. Specify sections and lengths and the essay problem disappears.
• The model filled your gaps.Every fact you don't provide is a fact the model may invent. Bracketed inputs make the gaps visible before you run the prompt.
• You accepted the first draft.Operating prompts improve with use. When an output annoys you, don't fix the output — add a rule to the prompt so it never happens again. After five runs the prompt is genuinely yours.
• One prompt tries to do three jobs. A review prompt that also writes content and plans pricing does all three badly. One prompt, one job, run separately.

Where to start

Don't collect prompts — run one. Take the weekly review above, gather your real numbers for ten minutes, and run it this Sunday. The output will be imperfect. Add one rule, run it again next week. By week four you'll have a review ritual that costs you fifteen minutes and misses nothing. If you want the broader system this habit slots into, read how one person runs a whole company on AI agents — the prompts are the hands, but the docs-first system is the skeleton. And the full prompt arsenal is there when two prompts stop being enough.