Prompt best practices

What actually improves prompt output quality in Aimogen Pro, with concrete before-and-after examples.

The default prompts shipped with Aimogen Pro are reasonable. Most quality complaints trace back to a prompt that has not been adapted to the site.

Be specific about the output#

The single biggest lever.

Weak:

Write an article about %%topic%%.

Better:

Write a 1,200-word practical guide to %%topic%% for small business owners with
no technical background.

Structure:
- Open with the specific problem this solves, in two sentences. No preamble.
- Four H2 sections, each with a concrete example
- A short conclusion with one clear next step

Constraints:
- British English
- Second person ("you"), active voice
- No statistics unless you are certain of them
- Do not use: "delve", "landscape", "in today's world", "it is important to note"
- HTML formatting with proper heading tags

Every line of the second version removes a decision the model would otherwise make badly.

Say what not to do#

Models have habits. Naming them is more effective than hoping.

Worth banning explicitly: opening with "In today's fast-paced world", the word "delve", closing with "In conclusion", meta-commentary about being an AI, and hedging every statement.

Give the model the context it needs#

A prompt that references %%post_content%% produces something grounded. One that does not produces generic text.

For Amazon rules, pass the actual product data and instruct the model to use only that. For YouTube rules, pass %%video_captions%%. For an editing prompt, pass the content being edited.

Use %%article_so_far%% in section prompts#

The most under-used placeholder. Without it, section four repeats section two, because the model has no idea what it already wrote.

Write section %%current_section%% of an article titled "%%post_title%%".

Already written:
%%article_so_far%%

Do not repeat anything covered above. Do not re-introduce the topic.

One instruction per line#

Models follow a list more reliably than a paragraph. Long prose instructions get partially ignored; enumerated constraints do not.

Put the instruction before the data#

Instructions first, then the material. When the material is long, the instructions at the top are what the model anchors on.

Ask for length in words, then enforce it#

Say "1,200 words" in the prompt and set AI Content Minimum Character Count on the rule. The prompt guides; the setting triggers a continuation if the output falls short.

Note that Max Tokens is a ceiling, not a target. Raising it does not make output longer. See Tokens, context and output length.

Use the global prepend and append fields#

House style belongs in Prepend Text To All Textual AI Prompts once, not in seven prompt fields:

Write in British English. Use active voice and second person. Do not use the
words: delve, landscape, realm, tapestry, journey. Do not open with
"In today's" or close with "In conclusion".

Iterate in the Playground#

Do not tune prompts by running bulk rules — that is slow and expensive. Use the AI Playground: change one thing, run, compare. Move the prompt into the rule when it works.

The Model Comparison tab runs one prompt across several models at once, which is the fastest way to find out whether a problem is the prompt or the model.

Add variation for bulk runs#

One fixed prompt across a hundred posts produces a hundred similar posts. See Spintax and dynamic text.

Match the model to the task#

A cheap model with a good prompt beats an expensive model with a vague one. But no prompt makes a small model reason well. See Selecting models.

Temperature#

Adjust temperature or top_p, not both.

Instruct against invention#

Models produce plausible statistics, dates and citations that are wrong. Where accuracy matters:

Do not include statistics, dates, prices or named sources unless they appear in
the material provided above. If you would need a fact you do not have, write
around it rather than inventing one.

That does not make output reliable, but it measurably reduces confident fabrication.

What to fix first#

When output is wrong, in order:

1. Which stage produced it? Change that prompt, not another one.
2. Is the prompt specific about structure and length?
3. Does it have the context it needs?
4. Are the model habits banned explicitly?
5. Is the temperature right?
6. Is the model capable enough?

Related#

• Placeholder reference
• Debugging prompts
• Content quality problems