How much detail is too much

The advice has a ceiling

Everything so far has pushed one direction: be more specific. That direction is right and it is where nearly all the gains are. But it does have a limit, and knowing where it sits will save you from the second-year mistake: the 600-word prompt with fourteen numbered rules that performs worse than the 120-word version.

Three things go wrong when a prompt gets too dense.

1. Constraints get silently dropped

You learned the mechanism two lessons ago: instruction-following is trained behaviour, not a checklist. In practice, past roughly five to eight hard constraints on a single output, models start satisfying most and ignoring some. Nothing announces the omission. The answer looks complete.

Test it and the pattern is easy to see: ask for a 100-word product description that mentions the warranty, avoids the word “quality”, includes the SKU, uses British spelling, ends with a question, contains no adjectives before “delivery”, and stays under three sentences. You will get a fluent paragraph missing two of those.

What to do instead:

• Rank them. “These three are non-negotiable; the rest are preferences.” Models honour a stated priority better than a flat list.
• Split the pass. Draft first, then a second prompt that checks the draft against the list and fixes violations. Checking is an easier task than composing under constraint, and it is separately verifiable.
• Make them checkable. “Under 150 words” you can verify at a glance. “Engaging but not salesy” you cannot, so it will drift and you will never quite know.

2. Contradictions get resolved without telling you

“Comprehensive but under 200 words.” “Warm and professional but completely neutral in tone.” “Cover every clause but only mention the ones that matter.”

Faced with a contradiction, a model picks a reading and continues fluently. It rarely says these two cannot both hold. You are left with an answer that satisfies half of what you asked and reads as if it satisfied all of it.

The fix is a single line you can add to any complicated prompt:

Before answering, list any instructions above that conflict with each other or that you cannot satisfy together. Then answer, following the ones I marked as required.

This costs a few seconds and turns a silent compromise into a visible one.

3. Detail that is not task-defining becomes noise

There is a real difference between two kinds of specificity.

Task-defining detail changes what a correct answer is. The reader is a school governor with ten minutes. The budget is 400,000 pesos. The export changed last week. Remove any of these and the right answer changes.

Decorative detail does not. The model’s supposed twenty years of experience. Three sentences on why this project matters to you. A preamble explaining that you value clarity and precision. Remove them and the right answer is identical.

Decorative detail is not neutral. It takes space in the context, and in a long prompt it dilutes the instructions that matter, because the model is spreading attention across everything present. A 400-word prompt where 250 words are throat-clearing is a worse prompt than the 150-word version.

The audit question for every sentence in your prompt: would a correct answer change if I deleted this? If not, delete it.

The exception worth knowing

Long prompts are entirely appropriate when the length is material rather than instruction. A 4,000-word prompt that is 3,800 words of pasted contract and 200 words of task is well-proportioned. A 4,000-word prompt that is 3,800 words of rules about how to behave is usually a system that should have been split into stages.

The ratio to watch is instructions to material. Most weak prompts are heavy on instruction and starved of material; a few advanced ones tip the other way.

Where the balance actually sits

For everyday work: one to three sentences of task, as much genuine material as the job needs, three to five constraints that you can check, one line for the output shape, one line for the escape hatch. That will be somewhere between 80 and 300 words plus whatever you pasted, and it covers the overwhelming majority of real tasks.

If your prompt is longer than that and not mostly material, the question to ask is not “what else should I add” but “which of these am I actually going to check”.