Common Prompt Engineering Mistakes and How Agencies Avoid Them

A model knows a lot about the world and nothing about your client. If you ask for “a product description for our new shoe,” it has to invent the audience, the selling points and the tone. The result is plausible and generic.

Context is the cheapest improvement available. State who the reader is, what the product is, what makes it different, where the text will appear and what the reader should do next. Paste source material such as a spec sheet or interview notes and tell the model to use only those facts. That last instruction also reduces invented claims.

A single prompt that says “research the topic, outline an article, write it in our voice, add SEO keywords, and suggest five social posts” usually gets a mediocre version of each task. Instructions compete for attention, and the model may drop some of them entirely.

Chaining works better. Break the job into steps and check each output before moving on:

• Step 1: summarise the source material and list key facts.
• Step 2: propose an outline; a person approves or edits it.
• Step 3: draft one section at a time from the approved outline.
• Step 4: edit for voice using the brand guide.
• Step 5: create derivatives, such as social posts, from the final text.

Each step is easier to test and fix, and a person stays in control at the points that matter.

Describing a tone in adjectives, such as “friendly but professional,” leaves room for interpretation. Showing two or three examples of real past content in that tone removes most of it. This is often called few-shot prompting.

Choose examples carefully. If all three examples start with a question, the model will start with a question too. Vary them in structure while keeping the voice consistent, and say what to copy (“match this tone and sentence length”) and what not to copy (“do not reuse these topics or phrases”).

Specific enough to check. “Keep it short” is not checkable; “under 120 words, two paragraphs, no bullet points” is. The same applies to style. Instead of “avoid clichés,” list the phrases you never want to see. Instead of “SEO-friendly,” name the target keyword and where it should appear.

Constraints worth stating explicitly

• Length in words or characters, with the platform’s real limit.
• Format: headings, lists, tables or plain prose.
• Words and claims that are not allowed.
• Reading level or audience expertise.
• Output structure when a tool will parse it, such as JSON with named fields.

Yes, and it is the mistake that costs agencies the most time. A prompt that works once may fail on the next input. Before a prompt becomes part of a regular workflow, run it on several different inputs, including awkward ones: a very short brief, a product with no clear benefit, a topic outside the usual niche.

Score the outputs against criteria agreed in advance, such as factual accuracy, voice match and edit time. Change one part of the prompt at a time so you know what made the difference. When a model update arrives, rerun the test set, because behaviour can shift between versions.

A few smaller errors appear often in teams new to structured prompting:

• Negative-only instructions: “don’t be formal” works less reliably than describing what to do instead.
• Overlong chats: earlier instructions lose weight, so restart with a clean prompt for new tasks.
• Hidden assumptions: abbreviations or internal terms the model cannot know.
• Asking for certainty: requesting “accurate statistics” without sources invites invention.
• No ownership: prompts scattered in personal chats instead of a shared, versioned library.

Most AI platforms let you set standing instructions that apply to every conversation: custom instructions in ChatGPT, projects with instructions in Claude, Gems in Gemini, or a system prompt when using an API. These are the right place for stable context such as brand voice, audience and banned phrases. The task-specific details then go in the individual message.

Templates help people who do not write prompts every day. A form with fields such as “audience,” “main message,” “proof point” and “format” produces a consistent prompt without anyone needing to remember the structure. The risk is that templates go stale. Review them when the brand guide changes, and keep a short note inside each one explaining why each instruction exists, so nobody deletes an important rule because it looks unnecessary.