Prompt Engineering Case Study: A Step-by-Step Walkthrough
The fastest way to understand prompt engineering isn’t a list of rules — it’s watching a single piece of content move from a vague first attempt to a genuinely useful draft. Below is an illustrative walkthrough, built from the kind of iteration cycle we run at Salterra, showing exactly what changes between a weak prompt and a strong one and why each change matters.
- Step One: The Naive First Prompt
- Step Two: Adding Business Context and Audience
- Step Three: Structuring the Output, Not Just the Content
- Step Four: Tone Calibration Against a Voice Brief
- Step Five: Generating the GBP Posts and Review Response
- Step Six: The Human Edit Pass
- What Changed Between Draft One and the Final Version
- Applying This Same Process to a Different Content Type
The scenario: a local HVAC company needs a service page targeting emergency furnace repair, plus three GBP posts and a review response, all in the client’s established voice. Nothing here is a real client’s data — it’s a composite example built to show the mechanics clearly.
Step One: The Naive First Prompt
Most people start with something like “write a page about emergency furnace repair for an HVAC company.” This produces exactly what you’d expect: generic, competent, forgettable copy that could belong to any HVAC business in any city. It hits the keyword, technically answers the topic, and says nothing that couldn’t be said by a competitor with the city name swapped out.
This isn’t a failure of the model — it’s a failure of the prompt. The instruction contained no business-specific detail, no audience context, and no constraint on structure or tone. Garbage in, generic out.
It’s worth pausing here because this is the exact point where most marketers give up on AI-assisted drafting and conclude “it just doesn’t sound right for us.” The conclusion is understandable but wrong — the naive prompt was never going to produce anything else. Diagnosing that the failure lives in the input, not the tool, is the mental shift that makes every later step in this walkthrough possible.
Step Two: Adding Business Context and Audience
The second pass adds real inputs: the business has served the metro area for over a decade, technicians carry same-day parts inventory, the target reader is often calling in a panic at 11pm with no heat and a cold house. The prompt now instructs the model to write for that anxious, time-pressured reader rather than a generic search visitor.
- Business detail: years in service, technician certifications, parts-on-truck policy.
- Audience state: stressed, cold, likely comparing three companies in the next ten minutes.
- Differentiator: guaranteed same-day dispatch window, stated explicitly rather than implied.
The output improves noticeably. It’s no longer interchangeable with a competitor’s page, because it now contains specific, verifiable claims instead of generic reassurance language.
Step Three: Structuring the Output, Not Just the Content
Good content also needs the right shape. The third revision adds explicit structural instructions: lead with the urgent-need answer in the first two sentences, follow with a short section on what causes a furnace to fail in cold weather, add a section addressing the “is this a safety issue” concern directly, and close with a clear, low-friction next step.
This step matters because unstructured prompts tend to produce well-written paragraphs in the wrong order — burying the answer the anxious reader needs three paragraphs down. Structuring the prompt itself, not just editing the output afterward, saves a full revision cycle.
Step Four: Tone Calibration Against a Voice Brief
With structure and content solved, tone is the last mile. The client’s voice brief specifies plainspoken, warm, never salesy, contractions encouraged, no exclamation points. Feeding that brief directly into the prompt as a standing instruction — rather than hoping the model infers it — closes the gap between “good generic copy” and “copy that sounds like this specific company.”
A useful technique here: include one or two short real examples of the client’s actual past communications (an email, a social caption) as style reference inside the prompt. The model imitates concrete examples far more reliably than adjective lists like “friendly and professional.”
Step Five: Generating the GBP Posts and Review Response
The GBP Posts
With the service page locked, the GBP posts come next, using a separate but related prompt. Each post needs a distinct angle rather than three near-duplicates: one built around the same-day guarantee, one around a seasonal maintenance tip, one around a specific technician’s tenure. The prompt explicitly instructs “these three posts must not repeat the same opening sentence structure” — a small but effective anti-sameness guardrail.
This step also enforces the character limit and a call-to-action variation requirement, since GBP posts read as spammy when every post ends with the identical phrase.
The Review Response
The review response prompt is the most tightly constrained of the batch. It takes the actual review text as input, includes an instruction to reference one specific detail from the review (not a generic “thank you for your feedback”), and caps length at two to three sentences. The prompt also enforces a hard rule: never make a promise or claim that wasn’t already true, since review responses are public and easily screenshotted.
The result reads as genuinely responsive rather than templated — the single biggest differentiator between review responses that build trust and ones that quietly signal a business is phoning it in.
Step Six: The Human Edit Pass
Every draft in this walkthrough still goes through a human editor before publishing. The editor’s job at this stage isn’t to rewrite from scratch — the prompt work already did the heavy lifting — but to verify factual claims, check that nothing sounds slightly off for the specific business, and tighten anything that reads as over-explained. This pass typically takes a fraction of the time a from-scratch edit would, which is the entire point of investing in the prompt upfront.
What Changed Between Draft One and the Final Version
Looking back across the six revisions, the improvements came almost entirely from adding specificity, structure, and real reference material to the prompt — not from asking the model to “make it better” or “make it more engaging,” instructions too vague to act on reliably. Each fix targeted one identifiable weakness: missing business detail, missing structure, missing tone anchor, missing anti-repetition guardrail.
That’s the transferable lesson. Diagnose the specific gap in a weak draft, then fix the prompt input that caused it, rather than repeatedly regenerating and hoping for a better result by chance.
Applying This Same Process to a Different Content Type
The mechanics in this walkthrough weren’t specific to HVAC or to service pages — they’re the same six moves that apply to almost any recurring marketing content type. Swap the scenario to a B2B software company drafting a comparison page, and the sequence looks nearly identical: a naive prompt produces generic feature-listing copy, adding real customer context and competitive positioning sharpens it, explicit structural instructions fix the order in which arguments land, and a voice brief with real examples of past marketing copy closes the final gap.
What changes between industries is the specific inputs — a plumbing company’s differentiator is same-day dispatch, a software company’s might be integration depth or implementation time — but the diagnostic process for improving the prompt stays constant. That consistency is exactly why it’s worth learning this process deliberately rather than treating each new content type as a fresh problem requiring its own trial and error from scratch.
Teams that internalize this six-step pattern typically cut their prompt development time in half on new content types, because they’re no longer guessing at what might be missing — they’re running a known checklist: business context, audience state, differentiator, structure, tone anchor, anti-repetition guardrail.
Frequently Asked Questions
How many revision passes does a typical prompt need before output is publish-ready?
In practice, three to five focused revisions — each fixing one specific gap like missing context, structure, or tone — usually gets a prompt to a reliably strong state for a given content type.
Should the human editor still review AI-assisted drafts if the prompt is well-built?
Yes. A strong prompt reduces editing time significantly, but a human should still verify factual claims and confirm the tone lands correctly for that specific business before anything publishes.
What's the single most impactful change in this walkthrough?
Adding real business-specific detail and audience context in step two produced the biggest quality jump — far more than any later tone or structure refinement.
Can this same process work for industries other than local services?
Yes. The underlying steps — add context, add structure, calibrate tone against real reference material, add anti-repetition guardrails — apply to B2B, ecommerce, and content-heavy sites just as well as local service businesses.
Is it better to write one long master prompt or several smaller task-specific prompts?
Smaller, task-specific prompts, as shown across the service page, GBP posts, and review response steps, are easier to test, debug, and reuse than one large prompt trying to handle every content type at once.
Terry has 30+ years in software and SEO. He’s the founder of Salterra Digital Services and SEO Spring Training, host of the Roundtable SEO Mastermind, and lead instructor at SEO University — teaching the exact tactics his team uses on client work.
Continue learning
Ready to master this?
This guide is one lesson from the Prompt Engineering course. Get every lesson, framework and checklist — plus the full 38-course catalog — inside SEO University.
Practitioner-focused training across the full digital marketing stack — from technical SEO to conversion optimization and the AI search era. By Salterra Digital Services, since 2011.