A vague request such as “write me a blog post about X” gives a model little information about audience, purpose, sources, or voice. The response may therefore be generic or unsuitable for publication.

The following techniques add context and constraints. They are starting points for evaluation, not a substitute for human editing or evidence.

Technique 1: The Style Sample

One way to communicate voice is to provide examples of writing you own or have permission to use. Remove confidential or personal information before pasting material into a third-party service.

The technique:

Before your main request, add this block:

“Here are three paragraphs from my existing writing. Please study the voice, sentence structure, and tone — and write the following in a style that matches:” [paste 3 short paragraphs]

Why it can help: A style sample gives the model concrete signals about sentence length, tone, and vocabulary. Review the response carefully because it may still miss the intended voice or copy phrases too closely.

What to avoid: Using samples from different contexts (formal email vs. casual blog) confuses the model. Use samples that match the context of what you’re about to request.

Technique 2: The Constraint Stack

Adding constraints isn’t limiting — it’s precision. Vague prompts produce vague output. Specific constraints produce specific output.

The technique:

Instead of: “Write an Instagram caption about morning routines”

Try: “Write an Instagram caption about morning routines. Under 150 words. Opens with a bold statement, not a question. Includes a ‘save this’ prompt mid-caption. Ends with a question about the reader’s routine. No motivational platitudes. No emojis in the first sentence.”

Why it can help: Constraints make rejection criteria visible. They do not guarantee a good result, so check whether the response follows them without distorting the meaning.

Common useful constraints:

  • Word count limits (“under X words” is more useful than “approximately X words”)
  • Opening restrictions (“do not start with…”)
  • Format prohibitions (“no bullet points in this section”)
  • Tone identifiers (“direct but not harsh, informed but not academic”)

Technique 3: The Counter-Opinion Injection

Some model responses hedge or present several sides without choosing an editorial angle. If the piece requires a point of view, supply your own and distinguish opinion from fact.

The technique:

Add a “My take:” line before asking for a draft:

“My take: Most advice about morning routines is useless because it ignores the fact that not everyone’s peak energy window is in the morning. My argument: you should build routines around your chronotype, not a 5am alarm clock.”

“Now write a blog introduction that opens with this perspective, not the typical ‘morning routines can transform your life’ framing.”

Why it can help: Stating the author's position gives the model an explicit argument to preserve. The author remains responsible for supporting factual premises and labeling opinion.

Important: Do not present a model-generated position as personal experience or belief. Decide the position yourself, disclose material AI assistance when appropriate, and verify its supporting facts.

Technique 4: The Negative Example

Telling the AI what not to do is often more powerful than telling it what to do.

The technique:

Run your prompt once. Get the output. When you see a specific phrase, opener, or pattern you hate, explicitly ban it.

“Rewrite this, but do not use the phrase ‘In today’s fast-paced world’ — or any variation of it. Don’t open with a rhetorical question. Don’t end with ‘By implementing these strategies…’”

Why it can help: Naming unwanted phrases or patterns makes a preference testable. The model may replace one cliché with another, so the output still needs a manual read.

Over time, build a personal “do not write” list specific to your content type.

Common offenders to ban:

  • “In today’s world…”
  • “Are you tired of…?”
  • “Let’s dive in”
  • “Game-changer” / “Leverage” / “Moving the needle”
  • “As we’ve seen in this article…”

Technique 5: The Iterative Sharpening Loop

A single response is only one draft. An iterative review can isolate a weak paragraph, unclear transition, or buried call to action without asking for a full rewrite each time.

The technique:

Think of the first output as a rough draft to edit, not a text to publish. Run a sharpening loop:

  1. First prompt: Get the full draft
  2. Second prompt: “The introduction is too slow. The hook needs to be more specific. Rewrite just the first three sentences.”
  3. Third prompt: “The third paragraph is the strongest. Expand it by 30% with a concrete example.”
  4. Fourth prompt: “Read the conclusion. The CTA is buried. Move the most important ask to the last sentence and cut the sentence before it.”

Why it can help: A narrow follow-up identifies the exact passage and editorial goal. Accept the rewrite only if it preserves meaning, evidence, and voice.

Instead of treating a response as ready or unusable, identify what can be revised and what must be discarded. Some outputs should not be used at all when their sources, rights, or factual basis cannot be verified.

Putting It Together

The techniques can be combined, but adding more instructions is not automatically better. Test whether the model follows each instruction and remove any that conflict or add noise.

AI-assisted content quality depends on several factors: the source material, prompt, model behavior, subject-matter knowledge, fact-checking, and human editing.

A maintained prompt library can document useful instructions and known failure modes, but it does not replace primary sources, original experience, or editorial accountability.