Why Shorter Prompts Make GPT‑5.6 Smarter: Insights from OpenAI’s Official Guide

OpenAI recently released an official prompt‑engineering guide for the new GPT‑5.6 model. The core recommendation is counter‑intuitive to older practices: make your prompts shorter . In internal coding‑agent benchmarks, removing duplicated instructions and unnecessary examples raised evaluation scores by roughly 10‑15% and slashed token consumption by 41‑66%, cutting overall cost by three to six‑tenths.

The guide explains why the old “add more context, add more safety clauses” habit no longer works. Earlier, weaker models needed repeated directives to understand intent; the newer model’s improved intent‑understanding treats repeated clauses as noise, causing distraction and even over‑cautious behavior (e.g., refusing legitimate actions because of over‑emphasized “never modify tests”).

Practical trimming workflow : start with a working prompt, delete one group of instructions or examples at a time, and re‑run the same evaluation. If the score does not drop, keep the deletion. The guiding principle is to state each instruction only once and describe each tool concisely.

The guide also provides a prompt skeleton with eight mandatory sections—role, personality, goal, success criteria, constraints, tools, output format, and stop rules—intended as a checklist rather than a verbose template. An example for a Go code‑review assistant illustrates the format:

角色：你是我们组 Go 后端仓库的 code review 助手。
性格：直接指出问题，不用客套。
目标：审查 PR 的改动，给出可执行的修改意见。
成功标准：每条意见指向具体文件和行号，说明原因。
约束：只看这次改动涉及的文件，别重构无关代码。
工具：可以跑测试和 lint，不能改代码。
输出：按严重程度列出问题，最多十条。
停止规则：信息不足时列出缺少的内容，别猜。

Two categories of content must never be removed: (1) examples that encode product requirements (e.g., mandatory output format) and (2) style guidelines that demonstrably improve scores. Everything else can be trimmed.

Another key shift is the move from “safety‑sentence stacking” (e.g., “never modify tests”) to a three‑tier action model:

Read‑only requests: the model may answer after reviewing material, but must not modify anything.

Safe modifications (tests, lint, builds): the model may act directly and run non‑destructive verification.

Potentially destructive actions (writes, external calls, spending): the model must first ask for confirmation.

Contradictory instructions (e.g., “never modify tests” + “fix all failing cases”) confuse the model and degrade performance more than omitting an instruction entirely.

Verification is essential because a generated answer does not guarantee task completion. The guide recommends giving the model built‑in verification tools (tests, type‑checkers, lint) and explicitly stating which checks are critical. For code changes, the model should run the most relevant tests, perform type checks, and, if verification is impossible, explain why and suggest the next manual check.

Finally, the guide stresses that prompt engineering is now about defining a contract: clearly state the desired outcome, boundaries, and proof of completion, letting the model decide the optimal internal steps. This contract‑style approach replaces the old “add‑more‑phrases” mindset and aligns with the higher cost of every unnecessary token in stronger models.