Skeleton-of-Thought Prompting: Guide with Examples (2026)

Last updated 2026-06-29 · Works with ChatGPT, Claude, Gemini
Quick Answer

Skeleton-of-Thought (SoT) prompting instructs the AI to first create a structural outline — the skeleton — and then expand each point into full content. This two-phase approach produces better-structured long-form output than asking for the final piece in one shot, because the AI commits to a logical structure before filling in the details.

What Is the Skeleton-of-Thought Framework?

Skeleton-of-Thought was introduced as a technique to improve both the structure and quality of long-form AI outputs. The key insight is that when asked to write a long piece directly, AI models often lose structural coherence midway through. By separating outlining from writing, the final piece has a logical arc the model committed to before expanding it.

Fields

T
Topic
What the final piece is about. Include the audience and purpose so the skeleton is shaped appropriately.
S
Skeleton points
The key sections, headings, or bullet points the AI should outline first. You can specify these yourself or ask the AI to generate them from the topic.
E
Expand instructions
How each skeleton point should be expanded — paragraph count, example requirements, depth of explanation.
O
Output format
The final formatting: Markdown, HTML, heading levels, word count, code block style.
+R
Rules
Constraints: Do not skip any section. Always show the skeleton before expanding.

When to Use Skeleton-of-Thought

Use Skeleton-of-Thought for any piece longer than 500 words where structure matters — guides, documentation, reports, blog posts, research summaries. It is particularly effective for technical writing where sections must logically build on each other. For short, single-topic outputs, the overhead of outlining is not worth it — use RTF or CRAFT instead.

Examples

Developer guide
Topic: A getting-started guide for developers integrating the Promptary REST API into a Node.js application.
Skeleton points: 1. Prerequisites 2. Getting your API key 3. Making your first request 4. Handling responses 5. Error handling 6. Rate limits
Expand instructions: Expand each section into 2–3 paragraphs. Include a working code snippet in sections 3, 4, and 5.
Output format: Markdown. H2 for each section. Code blocks with language tags. Max 800 words.
Rules: Do not skip any section. Show the skeleton first, then expand.
Blog post
Topic: Why developers should decouple AI prompts from application code — for a technical audience on DEV.to.
Skeleton points: 1. The problem: hardcoded prompts 2. What decoupling looks like 3. Real-world benefits 4. How to implement it 5. Conclusion
Expand instructions: Each section: 3–4 sentences minimum. Section 4 must include a before/after code example.
Output format: Plain text with H2 headings. Max 600 words. Conversational tone.

Compared to Other Frameworks

Skeleton-of-Thought vs RISEN: RISEN defines steps the AI follows to complete a task; Skeleton-of-Thought defines the structure of the output document itself. Skeleton-of-Thought vs Plan-and-Solve: Plan-and-Solve separates planning from execution for problem-solving; Skeleton-of-Thought separates outlining from writing for content creation. Skeleton-of-Thought vs CRAFT: CRAFT controls tone and format for shorter pieces; Skeleton-of-Thought is designed for longer pieces where structural integrity matters.

Save Skeleton-of-Thought prompts in Promptary — free

Save your first prompt in the structured editor and get a permanent REST API endpoint. Personal plan is free forever, no credit card required.

Start free

Frequently Asked Questions

What is Skeleton-of-Thought prompting?

Skeleton-of-Thought (SoT) prompting asks the AI to first generate a structural outline — the skeleton — and then expand each point into full content. The two-phase approach produces better-structured long-form output than asking for the complete piece in a single instruction.

When should I use Skeleton-of-Thought?

Use Skeleton-of-Thought for any output longer than 500 words where logical structure matters — documentation, blog posts, reports, and guides. For shorter outputs, the overhead of outlining is unnecessary.

Can I define the skeleton myself or does the AI generate it?

Both work. You can specify the exact sections in the Skeleton points field for tight control, or describe the topic and ask the AI to generate and confirm the outline before expanding. Specifying it yourself gives more predictable results.

Does Skeleton-of-Thought work with Claude, GPT-4, and Gemini?

Yes. The technique works across all major language models. The key is the explicit two-phase instruction: outline first, then expand.