“Write less, get more” is not a slogan — it is the actual mechanics of prompt design after GPT-5.6. The model’s intent understanding has improved to the point where it can infer your goal and how much effort a task deserves from context. At that point, step-by-step instructions stop being control and start being a constraint: you have pinned one specific path while the model could have picked a shorter one. OpenAI’s official prompting guidance is built entirely on this premise; this article distills it into a version you can apply directly.
Why Less Prompting Works Better
The directional evidence comes from OpenAI’s internal coding-agent evals — vendor-internal numbers, not independent verification: leaner system prompts improved scores by roughly 10–15% while cutting tokens by 41–66% and costs by 33–67%. Your product may not reproduce those figures, but the mechanism holds up: the same rule repeated three ways, tool descriptions irrelevant to the task, and examples that change nothing all spend the model’s attention budget.
What gets cut, though, is not the spec. Domain context, hard constraints, approval boundaries, and success criteria must stay; what you trim is steps and repetition, not acceptance standards.
Write Outcomes First, Conditions Second
The outcome-first skeleton is: goal, the context that matters, constraints, required evidence, success criteria, and output format. Pin the execution order only when the order itself is a business or safety requirement. This same skeleton is also the correct way to use Pro mode — you never need to write “use pro mode” or “think harder”; state the goal and the acceptance bar, and the model allocates reasoning depth on its own.
Trim One Group at a Time
The safest way to shorten a prompt is not a rewrite. Start from a version that already works, remove one group of instructions, examples, or tools at a time, and rerun the same evals. State each instruction once; keep only the examples that patch gaps you have actually observed; expose only the tools relevant to the task. Tone works the same way: instead of vague labels like friendly or empathetic, describe the concrete writing choices — how direct the answer should be, when to admit something cannot be done, whether a closing reassurance fits.
A Three-Tier Authorization Boundary
Autonomous work depends on stating authorization clearly, and three tiers are enough:
Read-only (answer, explain, review, diagnose, plan): read relevant data and report back; implement nothing.
Change (build, fix): complete in-scope local modifications and run non-destructive validation.
Confirm-first (external writes, deletes or overrides, purchases, scope expansion): get approval before acting.Explicitly listing safe local actions — reading files, viewing logs, modifying in-scope code, running tests — as directly executable is safer than scattering ask-first reminders across the prompt; repeated approval demands make the model pause even for normal reads. Keep the authorization policy in one place and the conflicts disappear.
Tool Routing and Stop Conditions
Tool routing is task-specific, and four things deserve explicit text: the bounded stage, the output schema, the limits on parallelism and retries, and a single handoff rule when two paths coexist. Every tool description should at least cover what it does, its important return fields, their types, and its failure behavior. Stop conditions work best as concrete formulas: “Stop when [condition]” and “Retry transient failures at most [R] times”. And when program_output is emitted separately from the final message, test both outputs; any resource-saving design still has to pass your existing evals.
When Programmatic Tool Calling Is the Right Call
Programmatic Tool Calling lets the model write JavaScript that calls tools and processes intermediate outputs inside a hosted runtime, and it shines in bounded, tool-dense stages: filtering, joining, ranking, dedup, aggregation, and validation — collapsing many large results into a small structure. It is ZDR-compatible and adds no extra container cost. Conversely, when a single direct call suffices, the next step depends on semantic judgment, an approval is required, or citations must be preserved, a direct tool call is usually the better choice.
Length, Tone, and Pro Mode
GPT-5.6 responds more concisely than GPT-5.5 by default, so during migration it is worth rechecking whether broad directives like “Be concise” still earn their place; when you need stable control over detail, use text.verbosity at low, medium, or high. The short-answer skeleton is conclusion, essential evidence, material caveat, next action — and the trimming order is introductions, repetition, generic reassurance, optional background. Pro mode is enabled on the same model via reasoning.mode: "pro"; do not switch to a separate Pro slug. Mode and effort are independent, both default to medium when omitted, and the cost is added latency plus extra tokens billed at standard rates within usage.
Migration Order and New Capabilities at a Glance
When upgrading, first keep your current reasoning effort as the baseline and test the same level and one level lower on representative tasks before deciding. Of the six levels — none, low, medium, high, xhigh, max — medium is the balanced starting point, low suits latency-sensitive workloads, and max is reserved for the hardest, quality-first jobs. On naming, gpt-5.6 aliases to the flagship gpt-5.6-sol, gpt-5.6-terra targets lower-cost high performance, and gpt-5.6-luna aims at high-volume traffic. The Responses API multi-agent beta coordinates parallel subagents inside a single instance and merges their results, cutting wall-clock time. Explicit prompt caching prices cache writes for reusable prefixes at 1.25x the uncached input rate with discounted reads, managed by prompt_cache_options.ttl in place of the older retention parameter. Persisted reasoning defaults to all_turns on GPT-5.6 and pairs with previous_response_id; with storage disabled or under ZDR, you resend encrypted reasoning items instead. On safety, instant classifiers for cyber and biology may block a request or pause mid-run for a few seconds of review, occasionally catching dual-use work by mistake; consumer-facing apps are advised to attach a safety_identifier so protections have a stable anchor.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.