Refining your prompts for GPT-5.6 Luna

This article applies to:Available in these QorusDocs Editions:

Enterprise+, Proposal Pro, QPilot Essential

The Default AI model is the model QorusDocs manages on your behalf for everything QPilot Assistant produces: Ask, QPilot Chat, Smart Skills and AI Text Smart Fields. QorusDocs keeps it on a current-generation model so you always benefit from the latest improvements in speed and accuracy without tracking model releases and retirements yourself. The AI Models article always lists the model currently behind Default.

When the Default model moves to a newer generation, every Smart Skill, AI Text Smart Field, Content Source Instruction and QPilot Custom Instruction you have saved carries over exactly as you wrote it. What can change is the shape of the output. Newer models follow instructions more precisely than older ones, and they tend to be more concise, so a prompt that gave you a half-page answer before may now give you a shorter one, or a more exhaustive one, depending on how it is worded. In most cases a small wording change brings the output back to what you want.

This article is for anyone who creates or maintains Smart Skills, AI Text Smart Fields or QPilot Custom Instructions. If you only run Smart Skills that someone else maintains, share this article with your Hub Admin or Designer.

In this guide you will learn what to expect, which prompts to look at first, how to review a prompt, and the adjustments that resolve the most common differences.

1. What to expect from a newer model
2. Which prompts to review first
3. How to review a prompt
4. Common adjustments
5. Where the temperature setting went
6. Best practices for prompts on the Default AI model
7. Getting help

What to expect from a newer model

Three things are true of current-generation models, and nearly every difference you notice comes from one of them.

• They follow your wording closely. If a prompt says “list all requirements”, you get all of them. If it says “keep it brief”, you get something brief. Older models filled in gaps with their own judgment; newer ones do what the prompt says.
• They are more concise by default. A prompt that relied on the older model’s natural padding may now return less. A prompt that asked for brevity may now return too little.
• They work best with lean prompts. Instructions repeated for emphasis, step-by-step scripts and “remember, this is important” phrases were ways of coaxing older models. Newer models do not need them, and too many of them can make the output worse rather than better.

You may also notice that QPilot Assistant is more careful about facts. Where an older model might have supplied a figure that was not in your document, the current model is more likely to say that the information is not specified. That is an improvement in accuracy, and the adjustments below show how to give the model room where you do want it to draw on general best practice.

Which prompts to review first

You do not need to review everything. Start with the prompts where the shape of the output matters most.

1. AI Text Smart Fields in PowerPoint templates. A placeholder on a slide does not grow to fit longer text, so these are the first to check. Run a test presentation and look for text that overflows or is cut off.
2. Smart Skills that quote, extract or outline. Requirement extractors, outline builders and compliance checks are the most sensitive to wording about quoting, completeness and structure.
3. AI Text Smart Fields in Word templates. The document reflows, so longer output is less disruptive, but a summary that doubles in length is still worth tightening.
4. QPilot Custom Instructions and Content Source Instructions. These apply to every output in your Hub or every question on a Content Source, so one loose rule here shows up everywhere.

How to review a prompt

1. Run the Smart Skill, Smart Field or Content Source question on a document you know well, ideally one you ran on the previous model, and compare the two outputs side by side.
2. Decide whether the difference matters. More detail, a different structure or a note that something is not specified are expected and often an improvement. Text that overflows a template, quotes where you wanted a summary, or a missing field you rely on, are worth fixing.
3. Open the prompt and find the line that produced the difference. It is usually an instruction with no bound (“all”, “every”, “detailed”, “where applicable”) or an instruction that was added to coax the older model (“be concise”, “think step by step”, the same rule stated twice).
4. Change only that line, using the adjustments below. Leave the rest of the prompt as it is, including any @ references to documents or Content Sources.
5. Run the prompt again on the same document and compare. One change at a time makes it clear what fixed the issue.
6. Save the prompt. If the Smart Skill is shared across your Hub, let the people who use it know what changed.

Common adjustments

Each adjustment below shows a line as it is often written, what the current model does with it, and a replacement. Use the replacement as a pattern and put in your own numbers.

The output is shorter, or a detail you rely on is missing

Blanket brevity instructions such as “be concise”, “keep it short” or “brief summary” now trim more than they used to. Replace them with what the short answer must include.

• Before: “Keep the summary brief.”
• After: “Lead with the conclusion, then include the deadline, the evaluation criteria and any mandatory requirements, each in one sentence. Omit background. Aim for 80 to 120 words.”

The output is longer or lists everything

Words like “all”, “every”, “comprehensive”, “detailed” and “thorough” are now taken literally. Replace them with a count.

• Before: “Provide a detailed summary of all requirements.”
• After: “Summarise the requirements in no more than 8 bullets of one sentence each. Group related requirements rather than listing each one.”

The output quotes the document instead of summarising it

“Verbatim quotes where applicable” or “quote the official language” now means quoting wherever possible. Ask for paraphrase in the body and give quotes their own place with a limit.

• Before: “Use bullet points and verbatim quotes (in quotation marks) where applicable.”
• After: “Use bullet points. In the body, paraphrase in your own words. Place verbatim quotes only under Verbatim Extracts, at most 3 per section, each no longer than one sentence.”

The structure goes deeper than you designed

An outline shown as 1., 1.1, 1.2 with no limit is read as a minimum, not a maximum. State the maximum depth.

• Add: “Use at most two levels of numbering (1., 1.1). Do not create a third level.”

The format is not what you asked for

“Use markdown formatting” or no format instruction at all lets the model choose, and it often chooses lists and tables. Name the elements you want and nothing else.

• Add: “One paragraph of plain prose. No bullets, tables or headings.” or “Bullets only; no tables.”

Check also that the prompt does not say one thing and show another. A prompt that says “no markdown” and then provides an output template written in markdown headings asks for two incompatible things, and the result will vary from run to run. Keep one or the other.

The output includes reasoning or explanations you did not ask for

“Justify each item”, “explain your reasoning” or “think step by step” now produce visible explanations. Where the reason matters, bound it to a reference or a phrase; otherwise remove the instruction.

• Before: “Justify each inferred section with reference to the RFP.”
• After: “Mark inferred sections (Inferred: RFP section X) with the section reference only.”

“Not specified” appears on every line, or the model will not give a figure it used to give

The current model does not invent what the document does not contain, and a rule like “state clearly when information is not found” is applied to every field. Scope the rule, and if you want best practice alongside document facts, say so.

• Before: “State clearly when information is not found.”
• After: “For the Deadline, Budget and Contact fields, write Not specified if absent. Omit any other field with no supporting content. Where the RFP is silent on approach, you may add general proposal best practice, marked (Best practice).”

Results vary from run to run

Look for two instructions that pull in different directions, such as “be concise” and “be comprehensive”, or “brief overview” and “include all details”. Older models quietly picked one. Choose the one you want and delete the other; do not add a third instruction to referee between them.

The prompt is long and the output is subtly off

Prompts grow over time. If yours contains the same rule stated more than once, phrases like “remember, this is very important”, or a role-play opener such as “You are an expert proposal writer”, try removing them. State each rule once, keep the context and constraints, and describe what a good result looks like. Remove one group of instructions at a time so you can see what each change does.

Text overflows a PowerPoint placeholder

Give the AI Text Smart Field a hard limit sized to the placeholder, and check the placeholder itself.

• Add: “Write no more than 45 words as a single paragraph. Do not use bullets or headings.”

If the limit is already tight and text still overflows, the placeholder may need to be larger or have autofit enabled in the template. See how to create AI-Powered Smart Templates.

A Smart Skill behaves differently from the same question in QPilot Chat

QPilot Custom Instructions apply to every output in your Hub, Content Source Instructions apply whenever that source is used, and the Smart Skill prompt sits on top. The current model honours all three fully, so a formatting or length rule at Hub level can override the Smart Skill. Keep tone, terminology and guardrails in Custom Instructions, and length and layout in the Smart Skill, and do not state the same rule in both places.

Where the temperature setting went

Current-generation models do not use a temperature setting, so it has been retired from Hub Settings and Smart Skills. If you used a high setting for varied, creative phrasing, or a low setting for repeatable results, say it in words instead.

• For variety: “Vary sentence structure and avoid repeating the same opening phrase across paragraphs.”
• For consistency: “Use the exact field names and order below in every response.”

Output on the current model is already more consistent from run to run than a low temperature setting gave you on older models, so most consistency rules are no longer needed.

Best practices for prompts on the Default AI model

1. State every bound as a number. Counts, word limits and depth limits remove the guesswork that older models filled in for you. “No more than 8 bullets” is clearer than “a short list”.
2. Describe the outcome, not the procedure. Tell the model what the finished output should contain and look like. Leave out the steps.
3. Say each thing once. Repetition and emphasis no longer help, and they can confuse priorities.
4. Separate facts from quotes. Ground every point in your document, but ask for paraphrase in the body and give quotes a named section with a limit.
5. Remove instructions that no longer earn their place. If removing a line changes nothing, it was a workaround for an older model.
6. Test on a document you know. One known document and one change at a time is the fastest way to see what each edit does.

Many prompts were refined over months by trial and error against an older model, so it is normal to find a few lines that are no longer needed. A shorter, clearer prompt is the usual result.

Getting help

If an output looks wrong rather than different, for example a figure that is not in your document, a referenced Content Source that was not used, or a Smart Field that comes back empty, contact helpdesk@qorusdocs.com with the prompt, where it runs, and the output. If you would like help adjusting a specific prompt, send the same details and our team will suggest an updated version.