Prompt engineering for SMBs: patterns that save real hours

The four habits

1. Tell it who it is. "You are an experienced HR generalist reviewing a job description for a 25-person construction company." One sentence of role + context changes the output more than any other trick.
2. Show, don't describe. Paste one or two real examples of what "good" looks like. Models are dramatically better at imitating examples than following abstract rules.
3. Ask for the format. "Reply as a 5-row table with columns Risk, Likelihood, Mitigation." Don't accept whatever shape the model decides to return - your downstream tools care.
4. Tell it what NOT to do. "Don't invent customer names. If the source doesn't say, write Unknown." This is the single biggest hallucination reducer for SMB use cases.

Template 1: Reusable prompt scaffold

Keep one of these per recurring task. Save them in a shared doc.

Role: <who the model is>
Goal: <what we want, in one sentence>
Context: <facts the model needs that aren't obvious>
Inputs: <paste the data here>
Format: <exact shape of the answer>
Constraints: <what to avoid, what to flag>

Template 2: The "draft for a human" pattern

For customer-facing work (emails, proposals, replies), never ship raw AI output. Use this:

Draft a reply to the email below in our voice (warm, concise,
no jargon). Mark with [VERIFY] anything you're not 95% sure about.
Mark with [ASK] anything that needs a decision from me before sending.

The [VERIFY] and [ASK] tags turn the model into a junior teammate instead of an overconfident intern. Adoption goes up sharply when people can trust the markers.

Template 3: Summarize without losing the facts

Summarize the document below in 5 bullets max.
Then list every number, date, name, and dollar amount mentioned.
If any of these would change a decision, flag it.

Common mistakes

• Asking too many things at once. One prompt, one job. Chain them.
• "Be more creative." Means nothing. Show an example of the tone you want.
• Not pasting the actual data. Models can't read your CRM. Paste the row.
• Trusting the first response. Re-prompt with "What did you miss?" - it often catches its own gap.

Where this gets you

Teams that internalize these habits cut their AI ramp time from months to days. The goal isn't to become an AI expert - it's to write briefs clearly enough that a model with no context can still do the job. That skill, it turns out, also makes you better at briefing humans.