Prompt Engineering for Claude 5 and GPT-5.6: Less Is More

Both Anthropic and OpenAI published prompting guidance in 2026 that overturns how most teams write prompts. The headline finding from both: leaner prompts score higher.

This post covers what changed and what it means for how you prompt modern models.

The OpenAI Finding

OpenAI's prompting guide for GPT-5.6 is blunt: state each instruction exactly once. In their testing, configurations with leaner system prompts improved evaluation scores by roughly 10 to 15 percent while reducing total tokens by 41 to 66 percent and cost by 33 to 67 percent.

This is a rare case where you can cut cost and improve quality simultaneously. The mechanism: GPT-5.6 is tuned to do more with less instruction. Repeated instructions, duplicated style notes, redundant examples, and process directions that do not change behavior all add noise that the model has to filter. Older models needed redundancy to obey instructions. Newer models interpret redundancy as conflicting signals.

The Anthropic Finding

Anthropic published the same result from a different angle. Their Context Engineering post showed that removing over 80 percent of Claude Code's system prompt caused zero regression on coding evals.

The model had better judgement than the prompt was encoding. The removed instructions were not helping - they were scaffolding for older models that needed more guidance.

The Shift from Rules to Judgement

The common thread: modern models have fundamentally different failure modes than their predecessors. Claude 5 and GPT-5.6 follow instructions carefully. The problem now is not insufficient instruction but conflicting instructions from different layers of the system prompt.

The shift is from prescriptive rules to decision rules. Instead of "always use X, never use Y," OpenAI recommends using "ALWAYS" and "NEVER" only for true invariants like safety rules. For judgment calls - when to search, ask, use a tool, or keep iterating - prefer decision rules that describe the destination rather than every step.

GPT-5.6 adds a verbosity parameter (low, medium, high) that replaces vague brevity instructions like "be concise" or "keep it short." These blanket rules can now make responses too brief because GPT-5.6 is already more concise by default than GPT-5.5.

The Minimal Viable Prompt Structure

Based on the guidance from both providers, we use a minimal prompt structure:

Identity and purpose (one sentence)
Key constraints (three to five items, only what actually matters)
Reference to external context (skills, tools, documentation)

Everything else lives in skills or tool descriptions loaded on demand.

OpenAI's recommended prompt structure follows the same shape: define the outcome, important constraints, available evidence, and completion bar. Leave room for the model to choose an efficient path.

Testing Prompts Like Code

Both providers emphasize the same workflow: prompt changes should go through the same review and testing process as code changes.

OpenAI recommends storing production prompts in application code, using typed inputs, code review, tests, and your normal deployment process to change model behavior. They are deprecating reusable prompt objects in the API.

Anthropic's approach is to remove one group of instructions at a time, then rerun the same evals. You cannot tell whether a behavior change came from the model, reasoning setting, prompt, tool set, or runtime if you rewrite everything at once.

Progressive Disclosure

For both Claude 5 and GPT-5.6, putting everything upfront wastes tokens and confuses the model. The pattern is progressive disclosure: load only what the current task needs. Skills are loaded by name, tools are loaded with minimal descriptions, and detailed instructions are retrieved just-in-time when the agent needs them.

The practical result: our agent system prompts are roughly 60 percent shorter than six months ago, and our eval scores are flat to slightly improved.

Read the full Context Engineering series: orioverse.com/blog/what-is-context-engineering

Prompt Engineering for Claude 5 and GPT-5.6: Less Is More