Prompting guide for GPT-5.5

In CODEXIS AI, a prompt does not just determine the shape of the answer. It also influences the scope of the research, the amount of source material processed, and usually the overall credit consumption as well. A well-phrased query therefore delivers a more accurate output, faster work, and greater cost predictability.

This guide summarizes practical principles that help you submit queries efficiently, professionally, and with better control over the result.

Why the wording of a query matters

CODEXIS AI does not work only with the final answer. Based on your input, it searches for relevant source material, evaluates the connections between them, and only then assembles the output. Credit consumption is therefore usually driven not only by the length of the answer, but above all by the breadth of the input, the number of sources examined, and the difficulty of the synthesis.

Put simply:

• the more precise the input, the less searching and usually the lower the credit consumption,
• the broader or more vague the input, the more source material the AI has to go through and the more likely the output will be expensive.

This matters especially for legal research, work with case law, and tasks that combine multiple legal regimes or multiple documents.

What a good prompt should contain

1. Role and procedural or business position

The AI should know from which perspective it is answering. The same legal rule is often assessed differently depending on whether you represent the contracting authority, the supplier, the plaintiff, the defendant, the employer, or the employee.

Example:

• "I am the contracting authority for a public procurement contract."
• "I am assessing the matter from the employer's position."
• "I need an argument for the recipient of a subsidy."

2. A brief factual framework

State only those facts that affect the legal conclusion. Without context, the AI has to imagine variants that may not be relevant to your case.

Example:

• "The supplier claims that the increase in the volume of performance can be carried out by amending the order. The original framework agreement sets a maximum volume and a fixed delivery deadline."

3. A precisely formulated question

Instead of a general topic, submit a specific question that the AI is to resolve.

Less suitable:

• "What about contract amendments?"

More suitable:

• "Is it possible under Section 222 of the Public Procurement Act to amend an order that has already been made so that the new delivery deadline also applies to the originally ordered performance?"

4. The required type of output

The AI should know exactly what it is supposed to return. The most common variants are:

• a brief yes/no conclusion with reasoning,
• research without case law,
• research with case law,
• a summary in 3 to 5 points,
• a draft argument,
• a draft letter or opinion for the client,
• an overview of risks and a recommended course of action.

Example:

• "I want a brief conclusion in 5 points."
• "Prepare a short research note including the 3 most relevant decisions."
• "Draft a basis for an internal partner decision, without unnecessary stylistic embellishments."

5. The required length

If you do not specify the length, the AI may produce a longer output than you actually need.

Example:

• "One A4 page at most."
• "Answer briefly, without long quotations."
• "First just the conclusion and a list of relevant provisions."

6. The scope of sources and the boundaries of the research

State where the AI should search and where it should not. This significantly improves the predictability of both the result and the credit consumption.

Example:

• "Work primarily with the attached contract."
• "Also use related provisions of Czech law, but without case law."
• "Include case law only if it is essential for the conclusion."

7. How to proceed with incomplete input

For more complex matters, it is worth adding an instruction on what the AI should do if decisive facts are missing.

Example:

• "If a reliable conclusion cannot be reached without additional factual circumstances, first ask 3 follow-up questions and do not carry out broad research."

Recommended prompt template

I am [role / position].

I am dealing with this case:
[brief factual framework]

I need:
- [type of output]
- [required length]
- [whether or not I want case law]

Work:
- primarily with [the attached document / a specific area of law / the content of CODEXIS]
- with case law [only if essential / only ÚOHS and administrative court decisions / only Supreme Administrative Court decisions from the last 3 years]

If the question cannot be answered reliably without additional factual circumstances,
first ask follow-up questions.

What to avoid in your input

The AI cannot be persuaded like a person. It can only be instructed correctly. If you put contradictory information, unclear instructions, or a pre-suggested conclusion into your input, the model tends to fill gaps with assumptions or to look for a compromise where none should exist.

Therefore, above all avoid these situations:

• input that is too broad, with no limitation on topic, time period, or type of sources,
• suggestive phrasing that pushes the AI toward a pre-selected result,
• vague instructions such as "tell me what is wrong" without specifying the area and purpose,
• contradictory instructions, for example when you simultaneously demand brevity, completeness, and broad case law.

How to reduce credit consumption without losing quality

Start narrow, then expand

For more complex tasks, it is advisable to split the work into two steps:

1. first map out the decisive legal questions,
2. only then develop one specific area in depth.

Example of the first step:

• "Identify a maximum of 5 legal questions that are decisive for this case. Without case law."

Example of the second step:

• "Develop only question No. 2. State the relevant provisions and the 3 most important decisions."

This approach tends to be cheaper and at the same time more manageable than immediately submitting a full research request.

Distinguish between document analysis and full research

Working on a specific document is usually more economical than broad research across the entire content of CODEXIS.

Less suitable:

• "Assess everything in this contract that could be legally problematic."

More suitable:

• "In the attached contract, check only the provisions on penalties, termination, and liability for damage."

State whether you require case law

Research with case law is typically more demanding than research without case law. If you only need an indicative legal framework, write so explicitly.

Example:

• "Answer without case law; the sections and a brief conclusion are enough for me."

Conversely, if case law is crucial, specify it as precisely as possible:

• "Find Supreme Administrative Court case law on this question from the last 3 years."
• "Focus only on ÚOHS and administrative court decisions."

Limit the output to what you actually need

A longer output does not automatically mean a better result. In practice, it is often more efficient to first request a brief conclusion and only then request an expansion.

Example:

• "First summarize the conclusion in 5 points. Elaborate the details only on the point you assess as the most risky."

Give the AI a head start

If you know the relevant legal regulation, section, or type of proceedings, state it. This narrows the AI's search space and usually gets it to the relevant result faster.

Example:

• "Address the question under the regime of Section 222 of the Public Procurement Act."
• "Assess it under Act No. 218/2000 Coll. and consider other regulations only subsidiarily."

Work iteratively

Do not take the first answer as final. If the output is not accurate or goes unnecessarily broad, it is usually more efficient to refine the prompt, add specifics, or have only a particular part of the conclusion developed, rather than submitting the entire research request again.

Example:

• "Elaborate only point No. 3 and add the relevant provisions."
• "Return to the conclusion, but this time assess the matter from the employer's point of view."

How to work with the model settings

In addition to the prompt itself, the result is also influenced by the chosen model settings. For everyday work, it is advisable to start from the default configuration and change it only when there is a specific reason to do so.

Detail

Lower detail is suitable for brief working answers. Higher detail makes sense when you want an elaborated output or a more detailed explanation.

Reasoning effort

Use higher reasoning effort mainly for genuinely complex questions, obscure topics, or situations where the AI has to connect multiple indirectly related sources. For precisely defined tasks, a lower or medium level is often sufficient.

Processing class

• Flex is suitable where immediate speed does not matter and the priority is a lower price. However, it may not always be available; under high load it is deprioritized and may return errors. It is best used in the morning.
• Standard is the universal default choice.
• Priority is suitable only when the priority is the fastest possible processing.

Context size

Standard context is suitable for ordinary conversations and most document tasks. Use extra large context only when working with very extensive source material or a long chat history, because it increases the cost.

For more detail on why it is advisable to keep the context as clean as possible in longer chats and to start a new conversation when branching, see Long conversations.

Practical examples

Example 1: a quick legal framework

Less suitable:

• "What about a premature petition under the Public Procurement Act?"

More suitable:

• "I am the legal representative of the petitioner. I need a brief assessment of whether, under the Public Procurement Act, a petition for review of the contracting authority's actions can be filed before a decision on objections. Answer without case law, in 5 points, and state the relevant provisions."

Example 2: working on a contract

Less suitable:

• "Go through the contract and tell me what's wrong."

More suitable:

• "Work only with the attached contract. Focus on the change to the order, the delivery deadlines, and the mechanism for increasing the volume of performance. Assess whether the described change may be problematic from the perspective of Section 222 of the Public Procurement Act. Divide the output into: (1) what follows from the contract, (2) legal risks, (3) recommended course of action."

Example 3: client output

Less suitable:

• "Write a reply to the client."

More suitable:

• "Based on this research, prepare a brief client summary in a professional tone. Max. 10 sentences. Do not include internal working notes or alternatives that are not relevant to the client."

The economic view

A properly formulated prompt does not only serve to improve the quality of the answer. It also helps to improve cost predictability.

According to the current model examples in the CODEXIS AI documentation, typical legal research without case law comes to 17,423 credits and case law research to 58,635 credits. Depending on the package size, this corresponds approximately to the following amounts:

In absolute koruna terms, even for more demanding tasks these are usually low costs compared to human research or analytical work. The main goal of a good input is therefore not to "make an expensive tool cheaper", but to achieve a more accurate result, better predictability, and fewer unnecessarily broad research requests.

Recommended workflow

If you are not sure how to phrase a query, the following procedure has proven effective:

1. briefly describe the role and the facts,
2. define the precise question,
3. specify the required type and length of output,
4. state whether the AI should work only with the document, or also with other sources,
5. decide whether case law is necessary,
6. only then, if needed, expand the input with another layer of detail,
7. if the first answer is not enough, refine the input or have only a specific part developed.

Final recommendation

You will achieve the best results when you work with CODEXIS AI in the same way as with a qualified external advisor: give it a clear brief, the necessary context, and specify exactly what output you expect from it.

Other related chapters:

• Uploading your own files
• Editable blocks
• CODEXIS AI - navigating the chat
• Model examples of tasks