Skip to main content
Back to Insights
• 13 min read

Advanced Prompting Techniques

Learn advanced AI reasoning frameworks including Step-Back Prompting, Self-Discover, Graph-of-Thoughts, and other advanced techniques that make frontier models 2-3× more effective on complex problems.

New to prompting? Start with Prompting Fundamentals: The GCSE Framework to build a strong foundation.

Technique Categories:

Complex Reasoning & Problem-Solving

Techniques for tackling difficult analytical challenges

Step-Back Prompting + Abstraction

Instructs the model to first identify the higher-level principle before answering. This technique improves responses on hard or novel problems by grounding responses in fundamental principles rather than surface-level pattern matching.

Example Prompt

## Goal
Recommend the best pricing model for our new premium analytics dashboard.

## Context
Current: $29/mo base tier. Competitors: $15–$99. Customers: mid-market companies (50–500 employees).

## Expectations
Step-Back First: What higher-level principle is this pricing decision really about?

Then Answer: Use only the step-back insight to give a concrete recommendation with exact price points.

Typical Output

The pricing question is really about capturing perceived value instead of cost-plus. Mid-market buyers pay for business insights and efficiency gains, not raw feature counts…

Recommended: $79/user/mo for premium tier (volume discounts at 100+ users).

Graph-of-Thoughts (GoT)

Helpful when you need explicit branching control over the reasoning process. This approach models thought as a directed graph rather than a simple tree, enabling more sophisticated problem-solving paths.

Example Prompt

## Goal
Find the real root cause of our 18% user churn spike last month.

## Expectations
1. Generate 8 diverse thought nodes (hypotheses).
2. Build the graph: show dependencies (→), merges (+), loops (↔), invalidations (X).
3. Execute the optimal path and give the final diagnosis.

Typical Output

N1: Pricing change → N4: Onboarding friction (+ merge) → N7: Competitor poaching (X invalidated)

Final traversal: N1+N4→N8

Root cause: New pricing tier confused existing users during onboarding.

Self-Discover

This technique has the model discover and compose its own optimal reasoning modules instead of you hard-coding Tree-of-Thoughts or Graph-of-Thoughts. This meta-cognitive approach allows the AI to dynamically select the best reasoning strategies for each unique problem.

Example Prompt

## Goal
Design a go-to-market plan for our new supply chain visibility platform.

## Expectations
SELF-DISCOVER protocol:
1. List 20 atomic reasoning modules humans use.
2. Select the 7 most relevant to this launch.
3. Compose them into a numbered reasoning plan.
4. Execute the plan step-by-step.

Typical Output

Selected modules: first-principles, analogies, counterexamples, probabilistic thinking…

Plan: 1. First-principles (what problem does supply chain visibility solve?) → 2. Analogies (Salesforce's early SMB strategy) → …

Final GTM: Start with logistics pilot program, then expand to manufacturing vertical…

Long-Form Efficiency & Structure

Optimized approaches for generating structured long-form content

Skeleton-of-Thought (SoT)

Generates a thin outline first, then expands every bullet in parallel in a second pass. This two-stage approach dramatically reduces generation time while maintaining content quality and coherence.

Example Prompt

## Goal
Write a 2,000-word whitepaper: "Digital Transformation Best Practices for Financial Services".

Stage 1 → Output ONLY a 12-bullet skeleton (no full sentences).

Stage 2 → Expand every bullet into a 150–250 word paragraph using subheadings.

Typical Output

Stage 1 Output
• Legacy System Assessment
• Cloud Migration Strategy
• Customer Experience Modernization
• Regulatory Compliance Framework
…

Stage 2 turns each bullet into full polished sections.

Accuracy & Reduced Hallucination

Systematic fact-checking for reliable, verifiable outputs

Chain-of-Verification (CoVe)

Significantly reduces hallucinations. Technique has the model plan verification questions, answer separately, then remove anything unverified. This systematic fact-checking approach aims to ensure every claim in the output can be substantiated with verifiable evidence.

Example Prompt

## Goal
Write a 400-word section: "Competitor analysis for Q4 product positioning".

## Expectations
1. List 10 specific, verifiable questions (e.g., pricing tiers, feature sets, market share).
2. Answer each in isolation. Mark [UNCERTAIN] if unsure.
3. Final section uses ONLY verified claims with [V#] citations.

Typical Output

Q1: Competitor A pricing? A: $49/user/mo [V1]
Q2: Market share? A: 23% in SMB segment [V2]
Q3: Latest funding round? A: [UNCERTAIN]

Final section contains only verified claims [V1][V2], omitting uncertain details.

Polish & Voice Enhancement

Iterative refinement and precise tone control

Self-Refine + Goal Gradient

Modern recursive polishing approach through iteration. The goal gradient ensures each revision demonstrates measurable quality gains, not just surface-level changes.

Example Prompt

## Current Draft
[Paste your draft here]

Score out of 100: Depth / Clarity / Originality / Punchiness.
Rewrite so every score rises ≥12 points (overall +15 min).
Output: NEW SCORES → REVISED DRAFT only.

Typical Output

Typical Output after 2 loops

NEW SCORES: Depth 94 | Clarity 96 | Originality 92 | Punchiness 95 (Overall 94)

[Polished version of content]

Directional Stimulus Prompting

Enhances styling and tone by requiring one or more exact sentences in the final output. This constraint-based approach leads the model to match your desired tone, style, and messaging, making it particularly helpful for brand consistency and editorial standards.

Example Prompt

## Goal
Write an 800-word piece on sustainable supply chain transformation.

## Constraints
Final paragraph MUST include: "The real opportunity isn't in carbon offsets, but in redesigning operations around circularity from day one."

Style: One concrete analogy per paragraph, zero clichés, measured optimism tone.

Typical Output

…The real opportunity isn't in carbon offsets, but in redesigning operations around circularity from day one.

Creative Synthesis & Ideation

Multi-perspective collaboration for comprehensive analysis

Storm / Multi-Perspective Synthesis

Uses a number of specialized personas to write independently, then a final persona takes a portion from each persona's writings. This collaborative approach combines diverse viewpoints while maintaining a cohesive final voice.

Example Prompt

## Goal
1,200-word piece: "Should we pursue a hybrid work model or return to office?"

Six perspectives (CFO, Head of HR, Team Lead, Remote Employee, Facilities Manager, Innovation Director) each write a 250-word take.

Then, as Executive Summary Author, synthesize into one balanced, actionable 1,200-word recommendation.

Typical Output

A single, seamless 1,200-word executive recommendation that balances all stakeholder concerns, yet reads as one unified strategic perspective.

Key Takeaway

These eight techniques naturally group into five challenge areas: complex reasoning and problem-solving, long-form efficiency and structure, accuracy with reduced hallucination, polish and voice enhancement, and creative synthesis and ideation.

Start with the one that fits your current task; that alone delivers the biggest immediate gain. Once comfortable, combine them freely. The strongest results come from mixing approaches to suit your specific workflow.

Sources & Further Reading

Explore the research and documentation behind these advanced reasoning frameworks: