Prompt Engineering

GPT-5.6 Prompting: Why Fewer Instructions Often Have a Greater Impact

The new OpenAI guidance challenges a common prompting practice: Longer sets of rules are not automatically better. Above all, GPT-5.6 needs a clear outcome, relevant boundaries, robust evidence, and a clear definition of “done.”

With the release of GPT-5.6 Sol, OpenAI has published prompting guidance that does not recommend adding even more prompt components. Instead, the focus is on the question: What information actually changes the model’s behavior—and what information merely makes the prompt longer, more contradictory, and more expensive?

OpenAI reports, based on a sample of internal coding agent evaluations, that more concise system prompts not only used significantly fewer tokens and incurred lower costs, but also produced better results. These figures are strictly indicative and must be verified through representative tests tailored to your specific use case.

1. Simplify first, then expand

The most important recommendation is this: Start with a prompt that already works, and remove only one group of rules, examples, or tools at a time. Then run the same evaluation cases again.

Typical candidates for reduction are:

• multiple versions of the same rule,
• Style or process guidelines that have no measurable effect on behavior,
• Examples that do not change the result,
• Specifications for behavior that the model already reliably exhibits,
• Tools and tool descriptions that are not relevant to the task.

However, the following must be retained: objectives, success criteria, security and business rules, evidence requirements, required tool paths, output format, and validation.

2. Describe the result—not every step

GPT-5.6 should be allowed to choose the most efficient route on its own. To do this, the prompt must describe the destination precisely:

• What should the user ultimately be able to see?
• What facts and sources need to be taken into account?
• What restrictions and permissions apply?
• What format is expected?
• When is the task complete—and when should you ask for clarification or stop?

Absolute terms such as “always” or “never” should be reserved for true invariants. When weighing options, decision criteria are more reliable than blanket prohibitions.

3. Inconsistencies are more dangerous than a lack of detailed rules

GPT-5 models treat prompts very much as a contract. If one instruction calls for maximum brevity, while another demands complete justifications, sources, alternatives, and next steps, instability is inevitable.

Before expanding the text, it’s therefore worth checking for potential conflicts: Which rule takes precedence in a conflict? What can be omitted? What content must be retained regardless of the desired length?

4. Define tone and collaboration separately

The guidance distinguishes between two levels:

• Personality: Tone, warmth, directness, formality, and writing style.
• Collaboration: When the model asks questions, makes assumptions, acts independently, explains uncertainty, and verifies results.

Both descriptions should be brief. They do not replace the objective or the success criteria. This distinction is particularly helpful for customer service representatives: A friendly demeanor does not necessarily indicate whether an agent is allowed to act independently.

5. Make autonomy and approvals explicit

Multi-stage AI systems need clear boundaries for their actions. A concise section can specify:

• For analysis and diagnostic tasks, read and report, but do not make any changes.
• For change requests, independently perform secure, local steps and appropriate tests.
• Require confirmation before external write operations, destructive actions, purchases, or significant increases in scope.

It is important to formulate these rules consistently from the start. Repeated requests to ask for clarification can cause unnecessary delays.

6. Provide only relevant tools

Tool descriptions should explain what the tool does, when it is used, which key fields it returns, and how errors are handled. A large toolbox is not a sign of quality if most of the tools are irrelevant to the task at hand.

Programmatic Tool Calling recommends OpenAI primarily for clearly defined processing steps such as filtering, sorting, deduplication, aggregation, or deterministic validation. For approvals, semantic judgments, citations, and final quality control, direct tool calls remain more suitable.

7. Research Requires an Evidence Budget

To ensure reliable answers, the prompt should specify which statements require sources and what happens when evidence is lacking. OpenAI recommends, in essence, starting with a targeted search and performing additional retrieval steps only if the core answer is still missing a relevant date, source, ID, or other required evidence.

This prevents two extremes: hasty guesswork and endless research that yields no new insights.

8. Don't instinctively increase the reasoning score

When transitioning to GPT-5.6, the previous reasoning effort should initially be retained as a starting point. Afterward, the same level and a lower level will be compared using real-world tasks. Higher levels are only worthwhile if evaluations show a relevant improvement in quality.

Before using more reasoning, you should check whether the prompt is perhaps missing only a success criterion, a dependency rule, a tool route, or a verification loop.

9. A Practical Approach to Structuring Prompts

For more complex tasks, the OpenAI structure can be broken down into seven short sections:

1. Role: Function and context of the model.
2. Personality: desired tone and communication style.
3. Goal: A visible result for the user.
4. Success criteria: Conditions that must be met before the answer is given.
5. Limits: Safety, Business, Evidence, and Side Effects.
6. Tools and Output: Selection Rules, Format, and Validation.
7. Stop Rules: Follow-up, Fallback, Repeat, or Abort.

Rule of thumb: Only add a statement if you can specify which observed misbehavior it is intended to correct.

10. Migration Without a Prompt Big Bang

OpenAI recommends a controlled process for migrating to GPT-5.6:

1. Switch models while retaining the reasoning effort used so far.
2. Perform representative evaluations before changing the prompt.
3. Remove outdated auxiliary structures, repetitions, and irrelevant tools.
4. Add only the smallest, targeted rule necessary to correct a measured deterioration.
5. Run the same evaluations again after every change.

If you change the model, prompt, toolset, and reasoning all at once, it will be nearly impossible to determine later which of these factors improved or worsened the behavior.

What Companies Should Take Away From This

Prompt engineering is maturing. The focus is shifting from linguistic tricks to a clear service agreement: desired outcome, verifiable quality, relevant data, permitted actions, appropriate tools, and a clear conclusion.

My conclusion: The best prompt for GPT-5.6 isn't the longest one. It's the shortest prompt that fully and consistently defines the goal, boundaries, and quality standards.

Source and starting point: OpenAI: Prompting guidance for GPT-5.6 Sol. The evaluation metrics cited are derived from an internal OpenAI sample and, according to the documentation, should be understood as indicative rather than a general guarantee.