Specificity Calibration

Your Prompts Are Either Too Vague or Too Controlling. Both Hurt.

Intermediate prompters often over-correct. They learn that vague prompts produce generic output, so they start writing extremely detailed prompts — specifying every structural element, listing every constraint, leaving nothing to the model's judgment. Then they're surprised when the output is stilted, mechanical, or misses the spirit of what they were after even while technically satisfying every stated requirement.

Over-specification is a real failure mode. Research on prompt specificity from late 2025 found that for open-ended reasoning and creative tasks, highly detailed prompts can actually constrain the model's ability to produce its best output by locking in a structure that predetermines the conclusion. The goal is not maximum specificity — it is correct specificity for the task type.

Under-Specification: What It Costs You

Under-specified prompts hand the interpretive work to the model. The model will make reasonable choices — but "reasonable" means statistically likely, not tailored to your specific situation. You get the average good response, not the response that fits your context.

The cost is proportional to how specific your requirements actually are. If you genuinely have no strong preference about tone, length, or structure, a brief prompt is fine — you'll get something useful. If you have clear preferences that you haven't stated, you will get output you have to significantly revise, and you will do that revision work instead of the AI.

Under-specified: "Summarize this article for me."

Calibrated: "Summarize this article in three sentences. Audience: a non-technical executive. Lead with the business implication, not the technical details."

Over-Specification: Where It Backfires

Over-specified prompts hand the interpretive work to you — before the task is done. When you enumerate every structural element ("first cover X, then Y, then Z, then a section on W, then conclude with Q"), you are essentially outlining the output yourself and asking the AI to fill it in. Sometimes that is exactly right. More often, you have locked in an outline you assumed was correct without knowing if it is the best structure for the content.

Over-specification also harms creative and analytical tasks specifically. When you tell the model exactly how to structure its analysis, you may be cutting off reasoning paths that would have produced a more insightful conclusion than your predetermined outline allowed.

Over-specified: "Write an analysis with exactly five sections: first, a background section covering the history of the technology; second, a section on current market dynamics with three bullet points; third, a SWOT analysis in a 2x2 format; fourth, a risks section with numbered items; fifth, a conclusion that recommends one of three options I will list."

Calibrated: "Analyze the current competitive dynamics in the enterprise CRM market. Audience: a VP of Sales deciding whether to switch platforms. Highlight the factors most relevant to a mid-size company with 200 salespeople. Length: 400–500 words."

The Calibration Framework

Use this framework to decide how much specificity a task needs:

• Always specify: Who the output is for, the format or medium, length or scope, and any constraints that would make a technically correct response wrong for your situation.
• Specify when you have a real preference: Tone, structure, what to emphasize, what to leave out.
• Leave open: How to arrive at a conclusion, the specific structure of analytical sections, the examples or evidence used (unless you have specific ones in mind), and anything where the model's judgment is likely better than your guess.

The Two-Pass Technique

When you are uncertain how much to specify, run two passes. First pass: minimal brief — just the task and audience. Read the output. What is it getting wrong? What choices did it make that you would have made differently? Second pass: add only the constraints that address the specific issues in the first output. This produces a targeted prompt rather than an exhaustive one, and it often surfaces constraints you would not have thought to include before seeing what the model's default was.

Task Type Guides Calibration Level

A rough guide by task type:

• High specificity appropriate: Business documents, templates, content with format requirements, output that must match a standard
• Medium specificity appropriate: Summaries, explanations, structured analysis, research synthesis
• Lower specificity appropriate: Creative work, brainstorming, open-ended exploration, tasks where you want to see what the model produces before deciding what you want

These are starting points, not rules. The pattern to internalize is: specify what constrains the output to be correct for your situation, and leave open anything where your specification would substitute your guesses for the model's judgment.