Tecof • September 14, 2026
AI Prompt Engineering 101: Getting Usable Output in E-commerce
In Brief
A prompt is the instruction you give an AI model; prompt engineering is the work of making that instruction repeatable and measurable. In e-commerce the difference-maker is not the model but the context you hand it. Asked to "describe this product", the same model writes something anyone could have written; given the product type, the audience, the brand voice, a banned-phrase list and a character limit, it writes copy that ships. As of 2026 the question is no longer "which model are we using" but "which job runs on which prompt, against which acceptance criteria".
Thursday afternoon, 14.40. Two hundred and forty products are going in for the new season and the descriptions are being generated. The first thirty come back like this: seventeen open with "made from quality materials", nine mention a feature the product does not have, four slip in a competitor's brand name. The copywriter starts fixing them, and fixing thirty takes longer than writing thirty from scratch.
The model is not the problem. Give that same model the product type, the audience, the tone, the length, a banned-word list and two examples, and twenty-six of the thirty ship untouched. The difference is the difference between a one-line request and a half-page contract.
What a Prompt Is and What It Means in E-commerce
A prompt is everything you feed the model: instruction, context, examples and any data. The model produces the most likely continuation of that input. So output quality depends far less on how clever the model is than on how determinative the input is.
The hiring analogy holds. Tell a new copywriter "write this product up" and the first draft probably will not sound like your brand. Give the same person the brand guide, the audience definition, two good examples and one bad one, and a "never say this" list, and the first draft lands close to publishable. AI works the same way, with one difference: you have to supply the guide every time, because the model does not remember the last job.
The five parts of a good prompt
- Role: who the model should behave like. "You are an e-commerce copywriter working in cosmetics, careful about regulatory claims."
- Context: product, audience, channel, brand voice. Missing context comes back as invented detail.
- Task: one clear job. "Write the description and suggest an SEO title and draft a social post" is three jobs and does better as three prompts.
- Format: the shape of the output. Length, heading structure, number of bullets, JSON or plain text.
- Constraints: what it must not do. Banned words, no health claims, no competitor names, no features you did not supply.
Writing all five takes a while the first time and no time at all across the next two hundred products. That is why a prompt is not a one-off sentence but an asset that goes into a library.
Zero-shot, one-shot, few-shot
Running a job with no examples (zero-shot) is fast and fine for generic work. One example (one-shot) fixes tone noticeably. Two to five examples (few-shot) is the cheapest way to make a model imitate your brand voice, and it is the single highest-return step in e-commerce. Choose examples that are representative rather than flawless: one short product, one long one, one difficult one. Adding a bad example tagged "do not do this" helps as much as adding a good one.
The Five Prompt Patterns That Earn Their Keep
The number of prompts that genuinely save time is small. These five cover most of a store's repeating writing work.
Product descriptions
Supply the product name, category, three standout features, the audience and a length limit. Ask for two paragraphs and three bullets. The critical constraint is this: "Use only the features given to you; add nothing that was not supplied." That one sentence closes most of the invented-feature problem. If product coding is messy the output will be too; a disciplined SKU structure is a prerequisite for bulk generation.
Category copy and SEO
Category page copy is a different job from product copy: it has to satisfy search intent. Put the target keyword, the secondary terms and the question competitors leave unanswered into the prompt. Write the 60-character meta title and 155-character meta description limits into the prompt itself; leave them out and the model writes long and Google truncates.
Review responses
Replying to negative reviews is where tone goes wrong most often. Give the prompt the review itself, the order type and your authority to resolve: "On shipping delays you may offer a reshipment; do not promise refunds." Leave the authority boundary out and the model promises something you cannot deliver, which costs more to undo than the review did.
Ad copy variants
Do not ask for one piece of copy; ask for five variants and make the model label the angle each one tests (price, speed, trust, social proof, scarcity). Then what you test is the angle, not the wording. We covered how to set up an A/B test to measure which one wins in a separate piece.
Data summaries and reports
Handing the model a raw table and saying "interpret this" is the weakest use there is. Ask the question yourself instead: "Find the three categories whose conversion rate fell in the last 30 days, give two plausible causes for each, and say which data to check next." Let the spreadsheet do the arithmetic; the model's job is interpretation, not maths.
| Pattern | Required input | Expected output | Most common failure |
|---|---|---|---|
| Product description | Name, category, 3 features, audience, length | 2 paragraphs + 3 bullets | Inventing features not supplied |
| Category and SEO | Target keyword, intent, character limits | Copy + meta title/description | Blowing past character limits |
| Review response | Review, order type, authority limit | Single-paragraph reply | Promising what you cannot deliver |
| Ad copy | Product, audience, channel, 5 angles | 5 variants + angle labels | Five copies of the same sentence |
| Data summary | Table, a precise question, period | Findings + next step | Making the model do arithmetic |
Measuring a Prompt: How Do You Check the Output?
Writing the prompt is the easy part. The hard part is deciding whether the output is good, because "I liked it" is not a measure: someone else will not like it tomorrow.
Writing acceptance criteria
For every prompt, write a five-item checklist and store it next to the prompt. For product descriptions: is it inside the character limit, does it contain a feature that was not supplied, does it use a banned word, does the target keyword appear once, does it sound like the brand. That list is the yardstick for both a human and an automated check, and it shows you what broke when you change the prompt.
Testing on twenty examples
Do not try a new prompt on two products and declare it working. Pick twenty products across easy, medium and hard, and run the prompt on all of them. Record the share that passes the acceptance criteria. A prompt below 70 percent does not go to production; the correction cost eats the gain. When the prompt changes, rerun the same twenty, so you notice when fixing one problem opened another.
| Stage | Measure | Threshold | Decision |
|---|---|---|---|
| First draft | Pass rate on acceptance criteria | Below 50% | Rewrite the prompt |
| Second round | Same rate | 50-70% | Add examples, add constraints |
| Production candidate | Same rate | 70-90% | Ship with review |
| Mature prompt | Same rate | Above 90% | Add to library, version it |
Do not read the rate as a number on its own. A prompt stuck at 85 percent can go live if it is clear who fixes the remaining 15 percent; if that is not clear, it cannot.
Prompting Agents: From Generating Text to Taking Action
The real change of the last two years is that a prompt no longer only produces text, it triggers action. Tell an AI agent to "pause ads for products whose stock fell below five" and the model does not write a sentence; it calls your store's functions and the ad platform's functions in sequence. That changes how prompts are written.
Write the goal, not the steps
In a text prompt you decide the steps. In an agent prompt you write the goal and the boundary and leave step selection to the agent: "Goal: stop ad spend on out-of-stock products before the campaign. Constraint: do not change budgets, only pause. When finished, list what you did." Spelling out every step turns the agent into a macro and removes the flexibility you were paying for.
Put the permission ceiling in the system, not the prompt
This is the most common mistake. "Never discount more than 15 percent" is a wish, not a control. The real boundary lives in the permissions on the key you gave the agent and in the function's own validation. A prompt can be talked round; a permission matrix cannot. We laid out that design in our guide to agentic commerce, and how functions are exposed to agents in the piece on APIs and integrations. On Tecof, more than seventy tools spanning CMS, catalog, orders and analytics can be opened to agents with defined permission boundaries.
KVKK and data minimisation
Putting personal data in a prompt means transferring that data to the model provider. Drafting a review reply needs neither the customer's name nor the order number; the review text and the order type are enough. Cover automated decision-making in your privacy notice, add the model provider to your list of data processors, and bring prompt and output logs into your retention policy. Any flow producing commercial messages must check consent in the system; a prompt cannot check it.
Building a Prompt Library in Thirty Days
Prompt work usually lives in one person's head and chat history. When that person leaves, the knowledge leaves. The timeline below turns scattered experiments into something the company owns.
Days 1-7: count the repeating work
How many product descriptions, category texts and review replies get written each week? How many minutes does each take? No prompts are written this week; numbers are collected. A job that does not consume five hours a week is not worth a prompt library.
Days 8-14: write prompts for the top three
Write all five parts, add two examples each, and put the acceptance criteria beside them. This week's output is three prompts and three checklists, not three hundred products.
Days 15-21: measure against twenty examples
Run each prompt on twenty real cases, record the pass rate, revise the prompt once, and rerun the same twenty. If the rate does not move, the problem is not the prompt but the missing data: description quality does not rise until product attributes are collected.
Days 22-30: version and hand over
Keep the prompts in one place with version numbers. Beside each one: what it does, the required inputs, the acceptance criteria, the last measured pass rate, and an owner. An ownerless prompt degrades within three months and nobody notices.
Here is the job for tomorrow morning: open five pieces of copy you had AI write and then edited by hand last week, and for each answer the question "what would I have had to say to avoid that edit?". The five sentences you end up with are the constraints section of your first prompt. On an e-commerce setup with AI built into the platform most of these patterns arrive ready-made and you only define the brand voice and the boundaries.
Frequently Asked Questions
Is prompt engineering a real job or a passing fad?
Its future as a separate job title is debatable; as a skill it is durable. As models improve, bad prompts increasingly still produce decent results, but the nature of the work is shifting from "generating text" to "tasking an agent", and there the ability to write a goal, a boundary and an acceptance criterion matters more, not less.
Is a longer prompt always better?
No. What helps is being determinative, not being long. Unnecessary detail distracts the model and raises cost. The test: if removing a sentence does not degrade the output, that sentence was unnecessary.
How do I stop the model inventing information?
The only reliable method is putting the information inside the prompt and adding the constraint "use only what is given; if something is missing, say it is missing." The second line of defence is a check that compares the output against the source. Saying "be accurate" is not a control.
Do I get better results prompting in English or in Turkish?
On current models the gap between Turkish and English instructions has narrowed sharply compared with two years ago. If you want Turkish output, give Turkish examples; what decides quality is the language and quality of the examples, not of the instruction.
The same prompt gives different results each time. Is that normal?
Yes, models are probabilistic. For consistency, add examples, tighten the format, and where possible ask for structured output (bullets, fields, JSON). Do not expect zero variance; target a pass rate against acceptance criteria instead.
Does AI-written content hurt SEO?
Search engines judge the value a page gives the reader, not how it was produced. What hurts is hundreds of pages cut from the same template that teach nobody anything. Add original data, real examples and a sound structure and the production method stops being the issue.
How many people and which tool should I start with?
One person and one document is enough. Before buying a dedicated tool for a prompt library, try versioning three prompts in a spreadsheet. The need for tooling appears once you pass about twenty prompts and more than one person starts editing the same one.
What is the real difference between an agent prompt and a normal one?
With a normal prompt you read and approve the output; with an agent prompt the agent acts and you see the result. That is why a goal, a boundary and a report-back ("list what you did") are mandatory in an agent prompt. And the permission is bounded by the system, not the prompt.
Who should own the prompt library?
The team that produces the content, not the technology team. The person measuring prompt quality and the person using the output should be the same; when they are not, the prompt goes unmaintained. Technical teams provide the tooling, not the ownership.