Prompting guide for GPT-5.6

This guide applies to GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna in CODEXIS AI. GPT-5.6 is better at understanding user intent and is more concise by default than GPT-5.5. You therefore do not need to prescribe every individual step. It is more important to define the goal, relevant materials, boundaries, and the required result clearly.

The prompt still affects the breadth of research, the amount of material processed, answer quality, and credit consumption. A good request is short but contains everything the model needs to produce a reliable result.

What is different with GPT-5.6

• Focus on the outcome. Describe what you want to achieve instead of scripting every step in detail.
• State each instruction once. Repetition, overlapping rules, and unnecessary examples can make a request worse and more expensive.
• Do not confuse brevity with incompleteness. GPT-5.6 is already concise by default. Specify which conclusions, evidence, caveats, and next steps it must not omit.
• Say when it should ask a question. Require a follow-up question when an ambiguity could change the legal conclusion or scope of work.
• Define authority for multi-step tasks. The model should know what it may do independently and what requires your confirmation.

What a good prompt should contain

1. The goal and a precise question

Start with the result you need. Replace a general topic with a specific question or the decision the answer is meant to support.

Less suitable:

• "What are the rules for amendments to contracts?"

More suitable:

• "Assess whether Section 222 of the Czech Public Procurement Act permits an existing order to be changed so that a new delivery period also applies to the originally ordered performance."

2. Your role and the relevant context

State the procedural or commercial position from which you are dealing with the matter and include only facts that may affect the conclusion.

Example:

• "I am a public contracting authority. The framework agreement sets a maximum volume and a fixed delivery period. The supplier proposes increasing the volume by changing an order that has already been issued."

3. Sources and required evidence

Define what CODEXIS AI should work with and how it should support its findings. This narrows the research and limits searches through irrelevant material.

Example:

• "Work primarily from the attached contract and Czech legislation. Add no more than 3 directly relevant decisions of the Czech Supreme Administrative Court from the last 3 years. Give the legal basis or source for each conclusion."

4. Constraints and boundaries

Say what is and is not part of the task. For agentic tasks, add which steps the AI may perform without further confirmation.

Example:

• "You may independently review the attached documents and CODEXIS content. Do not create or submit any filing and do not change any document without my confirmation. Do not address tax implications."

5. Success criteria and output format

Describe how you will know that the task is complete and what the result should look like.

Example:

• "Lead with an unambiguous conclusion. Then give the legal basis, no more than 3 material risks, and the recommended next step. Maximum length: one A4 page."

6. What to do when something is unclear

The model does not need to ask about a detail that cannot affect the result. Tell it when a question is genuinely necessary.

Example:

• "If a missing fact could change the legal conclusion, ask no more than 3 targeted questions. Otherwise continue and state your assumptions explicitly."

Recommended prompt template

I am [role / procedural or commercial position].

Goal:
[the decision or result I need]

Context:
[only the relevant facts]

Do the following:
- [precise question or task]
- [sources you should work with]
- [evidence or citations the output must include]

Boundaries:
- [what is outside the task]
- [what you may do independently and what requires confirmation]

Output:
- [structure and length]
- [what must not be omitted]
- [specific tone and intended audience]

If information that could change the conclusion is missing, ask no more than
[number] targeted questions. Otherwise continue and state your assumptions.

A shorter prompt is often better

GPT-5.6 usually does not need a long list of obvious steps. When revising a request:

• remove repeated instructions,
• combine rules that say the same thing,
• keep examples only when they define a required output,
• do not repeat general style guidance in several parts of the prompt,
• do not require a specific working process when only the result matters to you.

Less suitable:

• "Be concise. Do not repeat yourself. Write briefly. The answer must not be long. Summarize everything."

More suitable:

• "Answer in no more than 5 points. Preserve the conclusion, legal basis, material caveat, and recommended next step."

Controlling answer length and style

Set verbosity in the profile and content requirements in the prompt

The verbosity setting controls the usual answer length. Use the prompt for task-specific requirements such as the number of points, maximum length, or mandatory sections.

When asking for a short answer, define what must be preserved. For a legal output, this will usually be:

1. the conclusion,
2. the legal basis or controlling source,
3. a material uncertainty or caveat,
4. the recommended next step.

Describe the tone concretely

Instructions such as "write professionally" or "be empathetic" can be ambiguous on their own. State the intended reader and the specific writing rules.

Example:

• "Write for a client with no legal training. Lead with the conclusion, briefly explain legal terms, do not include internal working notes, and omit generic pleasantries."

Practical examples

A quick legal framework

• "I represent the applicant. Assess whether the Czech Public Procurement Act permits an application for review of the contracting authority's acts before a decision on objections. Use current Czech law, without case law. Answer in 5 points and preserve the conclusion, relevant provisions, material caveat, and next step."

Contract review

• "Work only with the attached contract. Review the clauses on penalties, termination, and liability for damage. For each finding, identify the specific clause, risk severity, and proposed amendment. Do not conduct general legal research. Return no more than the 5 most important findings in severity order."

An agent brief

• "Review the attached materials and relevant CODEXIS content, identify no more than 5 controlling legal questions, and propose the order for further research. Perform reading and analysis independently. Ask for confirmation before creating or changing a document. If a missing fact could change the order of priorities, ask about it first."

Saving credits without sacrificing quality

• Start with a precisely defined question and expand it only after the first answer.
• Specify the period, jurisdiction, source type, and maximum number of decisions.
• Distinguish between reviewing a specific document and conducting full research.
• First identify the controlling questions, then develop one of them in depth.
• In a long conversation, do not repeat the entire request; refer to the specific conclusion or section you want expanded.

For more information about costs, see Model examples of tasks.

Model settings and PRO mode

The same prompting principles apply to GPT-5.6 Sol, Terra, and Luna. Choose the model based on the required quality, speed, and cost; you do not need to rewrite the prompt for each variant. For ordinary work, start with the default profile and increase reasoning effort only for tasks that genuinely require deeper analysis.

For more on the available models and response profiles, see The model and its customization.

Checklist before sending

Before starting a demanding task, check:

1. Is the goal and precise question clear?
2. Does the request contain only relevant context?
3. Are the sources, jurisdiction, and period defined?
4. Does the AI know how to support its conclusions?
5. Are the boundaries and any approval requirements clear?
6. Are the required format, length, and intended reader specified?
7. Does the model know when to ask and when to continue with a stated assumption?

Source and related chapters

This guide adapts the recommendations in OpenAI — Using GPT-5.6 to CODEXIS AI.

• Long conversations
• Uploading your own files
• Editable blocks
• CODEXIS AI — navigating the chat