When you ask Copilot a complex question, it often generates a response from start to finish in a single pass. This can produce answers that are repetitive, incomplete, or poorly organized. Skeleton-of-Thought prompting is a technique that forces Copilot to first outline its answer and then expand each section. It solves the problem of shallow or sprawling responses by giving the AI a structured two-step process. This article explains what Skeleton-of-Thought prompting is, how it works, and provides concrete examples you can copy and adapt.
Key Takeaways: Skeleton-of-Thought Prompting
- Two-step prompt structure: Ask Copilot to generate a numbered outline first, then expand each point one by one.
- Reduces repetition and drift: The outline keeps the answer focused on the main topics you specified.
- Works best with long or multi-part questions: Use it for reports, comparisons, explanations, and decision-making tasks.
What Is Skeleton-of-Thought Prompting and Why It Works
Skeleton-of-Thought prompting is a method where you ask Copilot to first create a structured outline of its response before writing the full content. Instead of generating the final answer in one shot, the AI produces a skeleton of key sections and then fills in each section separately. This technique was originally studied in academic research on large language models to improve coherence and depth.
The core idea is that a single pass generation tends to produce linear text that may skip subtopics or repeat the same information. By separating planning from writing, Copilot can allocate more attention to each section. The outline acts as a roadmap that prevents the model from forgetting earlier points or going off-topic.
For business users, this translates into more structured emails, clearer reports, and more thorough analysis. The technique does not require special settings or plugins. You only need to modify how you write your prompt.
How the Two-Step Process Works
Step one: You ask Copilot to generate a list of sections or bullet points that cover the topic. Step two: You ask Copilot to expand each item from the outline into a full paragraph or section. You can do this in the same conversation by referencing the outline numbers.
This approach is different from asking Copilot to write everything at once. In a single-prompt scenario, Copilot must decide on the structure and content simultaneously. The skeleton method separates those two tasks, which improves output quality for longer answers.
How to Use Skeleton-of-Thought Prompting in Copilot
You can apply this technique in any Copilot interface: Microsoft 365 Chat, Copilot in Word, Copilot in Teams, or copilot.microsoft.com. The instructions below assume you are using the chat or compose pane.
- Open Copilot and start a new conversation
Click the Copilot icon in Microsoft 365 or go to copilot.microsoft.com. Clear any previous conversation to avoid context mixing. - Ask Copilot to create a skeleton outline
Type: “Generate a numbered outline for a report on [your topic]. Include at least five main sections with one subpoint each.” For example: “Generate a numbered outline for a report on remote work productivity. Include at least five main sections with one subpoint each.” - Review the outline and request expansion
After Copilot returns the outline, type: “Expand section 1 into a full paragraph. Use the subpoint as the main idea.” Repeat for each section. You can also say: “Now expand sections 2 through 5 one at a time.” - Combine the expanded sections into a final document
Copy each expanded section into a document or email. Copilot does not automatically merge them, so you must assemble the pieces yourself.
Example Prompt for a Business Analysis
Prompt: “Create a skeleton outline for an analysis of our Q3 sales decline. Include sections on market trends, competitor actions, internal process changes, customer feedback, and recommended next steps.”
After Copilot returns the outline, type: “Expand section 3 on internal process changes into a full paragraph.” Then continue with the remaining sections.
Common Mistakes and Limitations
Copilot Ignores the Outline Request and Writes Full Text
If Copilot does not produce a numbered outline, your prompt may be too vague. Rephrase to explicitly say “numbered outline” or “list of sections only.” Example: “List five sections for a report on customer churn. Do not write the full content yet.”
Expanded Sections Are Too Short or Repetitive
When expanding a section, add a constraint like “Write at least three sentences” or “Include one data point.” This prevents Copilot from producing a single sentence that adds no value.
Outline Sections Overlap in Content
If the outline has overlapping topics, ask Copilot to revise it before expanding. Type: “Revise the outline so that no two sections cover the same topic. Merge any duplicates.” This ensures each expanded paragraph adds unique information.
Skeleton-of-Thought Prompting vs Standard Prompting
| Item | Skeleton-of-Thought Prompting | Standard Single-Pass Prompting |
|---|---|---|
| Output structure | Predefined outline then separate sections | Single continuous text |
| Control over order | You choose which section to expand next | Copilot decides the flow |
| Best use case | Long reports, multi-part questions, comparisons | Short answers, simple definitions, quick lists |
| Risk of repetition | Low because each section is generated independently | Higher because the model may loop back to earlier points |
| Time to complete | More rounds of prompting | One round |
Use skeleton-of-thought when you need a thorough, well-organized answer. Use standard prompting when you need a quick response or the question is straightforward.
Conclusion
You now have a repeatable method to get structured, thorough answers from Copilot using skeleton-of-thought prompting. Start by asking for a numbered outline on your topic, then expand each section one at a time. For best results, add length and detail constraints to each expansion request. Try this technique next time you need a multi-section report or a complex analysis, and compare the output to a single-prompt answer.