Most small business owners in Canada don’t have a dedicated AI team. They have fifteen minutes between client calls, a ChatGPT or Claude subscription they’re paying for monthly, and results that feel inconsistent. One prompt gives a sharp, usable draft. The next one produces something generic enough to belong to any business on earth. The problem usually isn’t the tool – it’s that prompting without a structure is like cooking without a recipe. You might get something edible, but you won’t be able to repeat it. This framework is built specifically for Canadian SMB context: GST/HST considerations, bilingual markets, PIPEDA compliance language, seasonal patterns tied to Canadian fiscal calendars, and the reality that your “marketing department” might be you, at 9 p.m., on a Tuesday.
Why Generic Prompt Advice Doesn’t Quite Fit Canadian SMBs
Most prompt engineering guides are written for a US enterprise audience, or for developers building applications. The examples involve Silicon Valley startups, American tax codes, or assume you have a content team reviewing outputs. Canadian small businesses operate under different constraints.
You’re dealing with bilingual requirements if you serve Quebec or federal government clients. You’re working within PIPEDA (and provincial equivalents like Quebec’s Law 25) when handling customer data. Your fiscal year might end June 30th rather than December 31st. Your busy season might be tied to RRSP contribution deadlines in February, or the agricultural cycle in Saskatchewan, not Black Friday retail patterns.
Generic prompts produce generic outputs because they carry none of this context. The fix is systematic, not complicated. What we found surprising when testing this framework was how much output quality improved simply by adding two sentences of Canadian business context to otherwise ordinary prompts.
The 5-Step CRAFT Framework
The framework has five components, arranged as an acronym to make it easier to remember when you’re building prompts quickly. Each step corresponds to a piece of information the model needs to produce something actually useful.
- C – Context: Who you are, what your business does, where you operate
- R – Role: What persona or expertise you want the AI to take on
- A – Action: The specific task, as precisely as you can state it
- F – Format: How you want the output structured
- T – Tone & Constraints: Style requirements plus anything to avoid
None of these are new ideas individually. The value is in using all five together consistently, and in loading the Context step with genuinely Canadian-specific detail. Let’s walk through each one with real examples.
Step 1 and 2: Context and Role – Where Canadian Detail Actually Lives
Context is the step most people skip or thin out. A prompt that starts with “Write a newsletter for my accounting firm” gives the model almost nothing to work with. It will produce something plausible but hollow.
A better Context block looks like this:
Context: I run a two-person bookkeeping practice in Red Deer, Alberta.
We serve 40 small business clients, mostly in trades and light construction.
Our clients are owner-operators who file GST quarterly and mostly use
QuickBooks Online or Xero. RRSP season (January-March) and T4 filing
deadlines in February are our busiest periods. We send a monthly email
newsletter to clients.
Now the model knows your industry, geography, client profile, software stack, seasonal pressures, and communication channel. That’s eight distinct pieces of information that will shape every sentence it produces.
Role works best when it’s specific rather than flattering. Don’t say “You are an expert marketer.” Say:
Role: You are a plain-language business writer familiar with Canadian
small business accounting, CRA filing requirements, and writing for
tradespeople who prefer direct, no-jargon communication.
Specificity in the Role step reduces the chance of getting outputs that sound like they came from a generic MBA textbook. Our reading suggests that naming a concrete audience (“tradespeople who prefer direct communication”) has a bigger effect on tone than any explicit tone instruction alone.
Step 3: Action – Being Precise About What You Actually Want
The Action step is where most prompts underspecify. “Write a newsletter” is an action. But so is:
Action: Write a 250-word newsletter email reminding clients that Q3 GST
returns are due October 31st. Include one practical tip about reconciling
receipts before filing. End with a soft call to action inviting clients
to book a 20-minute check-in call if they have questions. Do not include
pricing information.
Notice what that Action block contains: word count, specific CRA deadline (October 31 for quarterly GST filers in Canada), content requirements, a structural endpoint, and a constraint. Each of those details removes a decision the model would otherwise make arbitrarily.
For Canadian SMB prompts, the Action step benefits from referencing real Canadian administrative landmarks:
- CRA deadlines: T4 slips by February 28th, T2 corporate returns 6 months after fiscal year-end, GST/HST returns by the last day of the month following the reporting period
- Provincial program references (Ontario Small Business Support Grant, BC Employer Health Tax, Alberta’s no provincial sales tax context)
- Bilingual requirements if the output needs French-language versions
- PIPEDA-compliant language for any customer-facing privacy notices
Being this precise feels like more work upfront. It genuinely is. But a prompt that takes 90 seconds to write carefully will often produce a draft that needs five minutes of editing. A vague prompt produces a draft that needs 25 minutes of rewriting – or gets discarded entirely.
Step 4 and 5: Format and Tone – Getting Outputs You Can Actually Use
Format instructions tell the model how to structure the output, not just what to say. This matters most when you’re going to paste the result directly into a tool (your email platform, a Google Doc, a social post scheduler) rather than treating it as a rough draft.
Format: Plain text only, no markdown. Use short paragraphs of 2-3 sentences.
Subject line first, then body, then sign-off. Total length 250 words maximum.
If you want HTML output for an email builder, say so. If you want a numbered list, say so. If you need the output in both English and French (common for federally regulated businesses or national-audience communications), say that explicitly:
Format: Provide the English version first, then a French translation below it,
separated by a horizontal line. The French should be professional Quebec
business French, not European French phrasing.
Tone and Constraints is the final step and often the most personally specific. This is where you encode your actual voice:
Tone: Warm but efficient. We don't use exclamation marks. We don't use
phrases like "exciting news" or "we're thrilled to announce." Write like
a knowledgeable neighbour, not a corporate newsletter.
Constraints: Do not mention competitors by name. Do not include any
statistics unless I've provided them in this prompt. Do not use the
word "leverage."
Constraints are underused. Telling the model what to avoid is often more effective than telling it what to do, because models will default to common patterns (corporate filler, vague statistics, buzzwords) unless you explicitly block them.
Putting It Together: A Full CRAFT Prompt Example
Here’s a complete prompt for a Canadian retail business using all five components:
Context: I own a women's clothing boutique in Saskatoon, Saskatchewan.
We've been operating for 8 years, have one part-time employee, and serve
women aged 35-60. We carry Canadian-made brands where possible and our
busiest season is September through November. We collect PST and GST
separately. Our Instagram has 2,100 followers.
Role: You are a social media copywriter who specializes in independent
retail and understands Canadian consumer culture, including sensitivity
around price (our customers are value-conscious but not bargain-hunters).
Action: Write 3 Instagram captions for posts celebrating a new fall
collection arrival. Each caption should mention that pieces are Canadian-made.
One caption should subtly address the value of buying Canadian versus
fast fashion imports. Captions should invite comments or saves, not
direct to "link in bio" every time.
Format: Three separate captions, numbered. Each 80-120 words.
Include 6-8 relevant hashtags per caption at the end.
No emojis in the caption body, only in the hashtag section if appropriate.
Tone: Genuine and grounded. This owner writes the way she talks -
direct, warm, no fluff. Never use phrases like "level up your wardrobe"
or "treat yourself." Avoid the word "stunning."
That prompt will produce three usable Instagram captions in a single generation. From our experience, a prompt at this level of detail cuts the typical revision cycle from three or four iterations down to one.
Maintaining a Prompt Library for Repeatable Results
The biggest efficiency gain from structured prompting isn’t the first time you write a CRAFT prompt. It’s the tenth time, when you’ve saved the Context and Role blocks as reusable components and only need to swap out the Action for each new task.
A simple prompt library can live in a plain text file, a Notion page, or even a pinned note on your phone. Structure it by use case:
- Customer emails (order confirmations, complaint responses, seasonal promotions)
- CRA and compliance reminders (deadline notices, document request follow-ups)
- Social media (by platform: LinkedIn copy differs meaningfully from Instagram)
- Quotes and proposals (executive summary language, scope descriptions)
- HR and internal communication (job postings, policy memos – keeping in mind provincial employment standards differ across BC, AB, ON, QC)
For Canadian businesses with French-language obligations – any federally regulated entity, businesses operating under the Charte de la langue française in Quebec – maintaining a separate French-language Context block saves significant time. French business tone differs enough from English that having a pre-written Role block describing the expected register (“professional Quebec business French, formal but not legalistic”) is worth building once and reusing consistently.
One important note on data handling: when using AI tools for prompts that involve customer information, be thoughtful about what you include. PIPEDA requires reasonable safeguards for personal information. Prompt text sent to cloud-based AI tools may be processed on servers outside Canada. Use anonymized or fictional stand-ins for actual customer data in your prompts, and review the data processing terms of whatever tool you’re using. The Office of the Privacy Commissioner of Canada has published guidance on AI and privacy worth reading if you handle sensitive client data regularly.
Common Prompt Mistakes Canadian SMBs Make (And Quick Fixes)
A few patterns come up repeatedly when small business owners share prompts that aren’t working:
- The context-free opener. “Write a blog post about tax tips.” Fix: Add your province, your client type, and the specific CRA context you want covered.
- The flattery role. “You are an expert in everything.” Fix: Name a specific domain, a specific audience, and a specific communication style.
- The open-ended format. Not specifying length, structure, or whether markdown is wanted. Fix: Add a Format block, even a short one. “Plain text, under 200 words, no bullet points” takes five seconds to type.
- Skipping constraints entirely. Without constraints, models default to common patterns – which means corporate-sounding language, vague calls to action, and filler phrases. Fix: List 3-5 things you don’t want.
- One-and-done prompting. Treating the first output as final rather than iterating. Fix: Build follow-up instructions into your library. “Revise the above to be 30% shorter and remove any sentence that starts with ‘As a…'” is a reusable edit instruction.
Structured prompting won’t make a mediocre AI tool good, but it will consistently get better results from a good one – and for a Canadian small business owner who’s doing their own marketing, operations, and customer service, consistency is the whole game.
– Auburn AI editorial, Calgary AB
Related Auburn AI Products
Building content or automations around AI? Auburn AI has production-tested kits: