Why Brand Rules Written as Instructions Outperform Brand Guides

A 1.3 billion parameter model trained to follow instructions was preferred by human raters over the 175 billion parameter GPT-3 base model. That finding — from Ouyang et al. at NeurIPS 2022 — is not an edge case. It is the defining result for how to write brand rules. Format closes more of the quality gap than scale.

The full citations are in the Instruction Following & Brand Rules research index.

InstructGPT: instruction format beats parameter count

Ouyang et al. (NeurIPS 2022) trained GPT-3 variants using reinforcement learning from human feedback (RLHF) to follow instructions. The trained models were labeled InstructGPT. Human raters were then asked to compare outputs from InstructGPT models against the 175B GPT-3 base model.

The 1.3B InstructGPT model was preferred. The mechanism: RLHF fine-tuning aligned the model's behavior with what humans actually asked for, independent of raw capacity. A 100× parameter disadvantage was overcome by training the model to respond to instructions.

For brand work, the implication is direct: how you structure brand rules is the primary quality lever, not which model your team uses. A smaller instruction-tuned model given clear, directive brand rules will more reliably follow them than a larger base model given the same rules in descriptive prose. The instruction format matters.

FLAN: training format determines generalization

Wei et al. (ICLR 2022) fine-tuned a 137B model — FLAN — on over 60 NLP tasks, each presented as a natural language instruction. They then tested it zero-shot (no task examples provided) on held-out tasks it had not seen during training.

FLAN surpassed the zero-shot performance of the 175B GPT-3 on 20 of 25 evaluated datasets. The mechanism was instruction format: by training on tasks framed as explicit natural language instructions, the model learned to generalize that format to new tasks.

The practical translation for brand rules: "always use X" and "never use Y" are instruction-format directives. "Maintain an approachable but professional tone" is descriptive language that requires the model to interpret what "approachable" and "professional" mean in context before it can act. The research says directive format produces more reliable generalization. Descriptive language produces more inconsistency.

A specific, testable rule — "keep paragraphs to three sentences maximum," "do not use 'leverage' as a verb" — is actionable. The model can verify its own compliance without interpretation. A rule that describes a quality — "sound like a trusted expert" — cannot be verified without judgment, and judgment produces inconsistency.

Constitutional AI: compact principles enable self-correction

Bai et al. (Anthropic, 2022) introduced Constitutional AI: rather than requiring human labels on every undesired output, the researchers provided an LLM with a written constitution — a small set of named principles — and had it self-critique and revise its outputs toward alignment.

The mechanism is called RLAIF: reinforcement learning from AI feedback. The model uses the constitution to generate its own preference feedback, reducing the need for human labeling of each individual failure.

The marketing application of this structural insight: a compact, named set of brand principles at the session level provides the frame the model uses to evaluate its own outputs. Rather than correcting every draft, the team encodes the principles and the model applies them continuously. Fewer principles that are specific produce more consistent self-correction than many overlapping principles that require priority judgment.

Constitutional AI is the research basis for a brand constitution approach: not a comprehensive brand guide, but a small set of specific, named, directive principles that can be stated, loaded, and consistently applied.

What actionable means

The instruction-following literature converges on one distinction: actionable directives vs. quality descriptors.

An actionable directive specifies a behavior: "Begin every section with the core claim." "Do not use rhetorical questions as headings." "Name the specific evidence before the conclusion." These can be evaluated for compliance.

A quality descriptor specifies an outcome: "Sound credible." "Be clear and concise." "Match the brand voice." These require the model to interpret what the descriptor means before acting, and that interpretation step introduces inconsistency.

Writing brand rules as actionable directives — and keeping the set compact enough to load consistently — is the implementation of the instruction-following research. The research behind it is in Instruction Following & Brand Rules.

For the practical application to briefing and context structure, see how to brief an AI and how context, memory, and prompts differ.

Frequently Asked Questions

Does the choice of AI model affect how well it follows brand rules?

Yes, but the InstructGPT finding (Ouyang et al. 2022) shows that RLHF training — not raw parameter count — is the primary predictor of instruction-following reliability. A smaller instruction-tuned model given clear brand rules may outperform a larger base model given the same rules. For brand tasks, the question is whether the model has been trained to follow instructions, not how large it is.

How many brand rules can an AI system follow consistently?

The research does not establish a specific number, but the instruction-following literature suggests that generalization depends on instruction clarity and specificity, not volume. Vague or overlapping rules create conflicts the model must resolve — and resolves inconsistently. The Constitutional AI research (Bai et al. 2022) found that a compact, named set of principles produces more consistent self-correction than a comprehensive list of constraints. Fewer, specific rules outperform more, descriptive rules.

Why do AI models sometimes ignore brand rules even when they're in the prompt?

Several factors contribute. Rules buried in a large context block are less reliably followed (Liu et al. 2023, Lost in the Middle). Rules written as quality descriptors require interpretation that the model resolves inconsistently. And some rules conflict with patterns the model has strongly internalized from training — a pattern that can override stated instructions when the rule is not specific enough. The structural fix is shorter rules, instruction format, placed at the top of context.

Is there a limit to how specific brand rules should be?

The risk of over-specification is rules that are too narrow to cover the actual range of situations the model encounters, producing correct behavior in the specified case and unconstrained behavior everywhere else. The Constitutional AI approach addresses this by encoding principles that generalize — rules at the right level of abstraction to apply across contexts without requiring a separate rule for every scenario. Specific enough to be actionable; abstract enough to cover the class of situations, not just the one case.