Past ~7-10 core directives, instructions compete for attention; quality plateaus or regresses while token cost climbs linearly.
Imagine handing a new hire a sticky note with three rules. They follow them. Now hand them a sticky note with thirty rules. They read the first few, skim the middle, and quietly forget half by lunchtime. Prompts work the same way. The model has a finite amount of attention to spread across instructions. Once you stack too many, some directives get ignored, others get half followed, and the behavior turns inconsistent. The fix is not more instructions. The fix is fewer, sharper instructions that each carry real weight.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Over-instruction is the failure mode where a team keeps adding directives to a system prompt and watches compliance get worse, not better. It is one of the most common quality regressions in production prompt engineering, and the instinct that causes it is exactly the wrong one: when the model misbehaves, add another rule.
The argument in this deep dive is that prompts have a directive budget, that the budget is set by attention competition rather than context window, and that the practical fix is a consolidation pass rather than a longer rule list. The interview-worthy framing is to name the budget, name the mechanism, and describe the rewrite pass that gets the prompt back inside it.
The sections below walk through the mechanism, the diagnostic, the rewrite pattern, and the production failure modes a senior engineer should be able to call out from a prompt review.
Why compliance is not linear in directive count
The intuition behind adding rules is that each rule covers one more case, so more rules cover more cases. This holds for the first few rules. It stops holding around seven to ten.
The reason is that every directive has to be attended to during generation. The model is doing a single forward pass over the prompt and weighting tokens against each other. When the system prompt has four directives, each one gets a healthy share of attention weight and the model follows it. When the same surface has twenty directives, the weight per directive thins out, and the model starts attending strongly to some while half ignoring others.
The failure is not uniform. Some directives keep their grip because they are positioned at the top, repeated, or paired with a worked example. Others quietly degrade. The team sees inconsistent behavior and adds another rule to fix it, which thins the budget further. The loop is self reinforcing if the diagnostic is not applied.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Anthropic's prompting guide for Claude Opus 4.7 warns that piling on constraints often degrades compliance and recommends consolidation plus positive framing.
- OpenAI's system prompt advice for GPT-5.5 nudges teams toward fewer, higher-leverage rules with worked examples rather than long rule lists.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you measure which directives are actually being followed?
Per-directive eval slice with a small held-out set per rule; track compliance rate per directive, not just aggregate; flag any directive under eighty percent for consolidation.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Treating compliance as a linear function of instruction count, so every new edge case becomes another bullet point in the system prompt.
60 second bullets to scan on the way to the call.
Why compliance plateaus past 7-10 directives
Attention competition as the underlying mechanism
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.