Opus Extended Thinking

Extended & Adaptive Thinking

This skill covers Claude's ability to reason deeply before responding — both the prompt patterns that elicit genuine deliberation (model-agnostic, below) and the current thinking API (how to turn it on and tune its depth). It applies to the current Claude family: Fable 5 (most capable), Opus 4.8 (default), Sonnet 4.6, and Haiku 4.5.

Model IDs and API details move fast. This skill is dated; always confirm against platform.claude.com.

How thinking works today (the API)

The modern control is adaptive thinking plus the effort parameter — not a fixed token budget. The older thinking: {type: "enabled", budget_tokens: N} is deprecated on Opus 4.6 and removed (returns a 400) on Opus 4.7, Opus 4.8, and Fable 5 — do not use it on current models.

# Current pattern (Opus 4.8 / 4.7 / 4.6, Sonnet 4.6)
response = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=16000,
    thinking={"type": "adaptive", "display": "summarized"},  # default display is "omitted"
    output_config={"effort": "high"},  # low | medium | high | xhigh | max
    messages=[{"role": "user", "content": "Solve this step by step..."}],
)

• effort controls depth and overall token spend. Use xhigh for coding/agentic work and a minimum of high for most intelligence-sensitive tasks; drop to medium/low for routine work. max is for the hardest, latency-insensitive cases (max/xhigh are Opus-tier + Fable 5, not Haiku).
• display defaults to "omitted" on Fable 5 / Opus 4.8 / 4.7 (thinking still happens and is billed; the text is just empty). Set "summarized" to surface a readable summary to users.
• Fable 5: thinking is always on — omit the thinking parameter entirely (an explicit {type: "disabled"} 400s), and the raw chain of thought is never returned (only summaries).
• Stream any request with a large max_tokens (>~16K) to avoid HTTP timeouts.

Everything below is about eliciting good deliberation through prompting — it works regardless of which model or effort level you choose.

Understanding deliberate reasoning

What it is

Adaptive/extended thinking lets the model:

• Deliberate longer before producing output
• Consider multiple perspectives and approaches
• Self-critique and refine reasoning
• Handle complex multi-step problems
• Produce more thoughtful, less reactive responses

When it engages

With adaptive thinking on, the model decides when and how much to think — typically when:

• Problems require multi-step reasoning
• Questions have nuance or ambiguity
• Tasks require weighing tradeoffs
• Complex analysis or synthesis is needed
• Creative work benefits from deliberation

Prompt Patterns for Deep Thinking

Pattern 1: Explicit Deliberation Request

Before answering, take time to:
1. Consider the problem from multiple angles
2. Identify potential pitfalls or edge cases
3. Weigh different approaches
4. Then provide your best response

Question: [Your complex question]

Pattern 2: Multi-Perspective Analysis

I need a thorough analysis of [topic].

Please consider:
- The technical perspective
- The business/practical perspective
- The user/human perspective
- Potential risks and downsides
- Long-term implications

Take your time to reason through each before synthesizing.

Pattern 3: Step-by-Step Reasoning

Walk me through your reasoning on this problem step by step.
Don't jump to conclusions - show your work.

Problem: [Complex problem]

For each step, explain:
- What you're considering
- Why you're making that choice
- What alternatives you rejected and why

Pattern 4: Devil's Advocate

I'm planning to [action/decision].

Before you help me execute this:
1. Steelman the opposing view - what's the best argument against?
2. What am I potentially missing or underweighting?
3. What would make this fail?
4. Only then, if it still seems right, help me proceed.

Pattern 5: Synthesis Request

I have [number] different sources/perspectives on [topic]:
[List them]

Please synthesize these into a coherent understanding that:
- Identifies where they agree
- Explains where they conflict and why
- Weighs the evidence/arguments
- Provides your reasoned conclusion

Task Types That Benefit Most

1. Strategic Planning

Prompt pattern:
"Help me think through [strategic decision].
Consider: current state, desired outcome, constraints,
multiple paths, risks of each, second-order effects.
Don't rush to a recommendation - reason through thoroughly first."

2. Code Architecture Decisions

Prompt pattern:
"I'm deciding between [Option A] and [Option B] for [system].
Before recommending, analyze:
- Scalability implications
- Maintenance burden
- Team capability fit
- Migration complexity
- What we might regret in 2 years"

3. Creative Development

Prompt pattern:
"I want to create [creative work] about [theme].
Before writing, explore:
- Different narrative approaches
- Tonal variations
- Structural options
- What makes this theme resonate
Then develop the direction that feels most promising."

4. Problem Diagnosis

Prompt pattern:
"[System/situation] is exhibiting [problem].
Don't jump to the obvious cause.
Consider:
- What are all possible causes?
- What evidence supports/refutes each?
- What's the most likely root cause?
- What would I need to verify to be sure?"

5. Research Synthesis

Prompt pattern:
"Help me understand [complex topic].
I need you to:
- Break down the key concepts
- Explain how they relate
- Identify what's well-established vs debated
- Note what's commonly misunderstood
- Synthesize into a coherent mental model"

Anti-Patterns to Avoid

❌ Rushing to Action

Bad: "Write me a function that does X"
Better: "I need a function that does X. Before coding,
        let's consider: edge cases, error handling,
        interface design, testing approach."

❌ Accepting First Answer

Bad: "What's the best database for my app?"
Better: "What database should I use for [specific requirements]?
        Consider multiple options, tradeoffs between them,
        and what factors should weight most heavily."

❌ Binary Thinking

Bad: "Should I do A or B?"
Better: "I'm choosing between A and B, but I'm open to
        alternatives I haven't considered. What's the
        best path forward given [context]?"

❌ Ignoring Uncertainty

Bad: "Tell me the answer to [complex question]"
Better: "Help me understand [complex question].
        Be explicit about what's certain vs uncertain,
        and what evidence supports different views."

Extended Thinking Triggers

Use these phrases to encourage deeper deliberation:

Analytical Triggers:

• "Think through this carefully..."
• "Consider all angles before responding..."
• "What are the non-obvious implications?"
• "What am I likely missing?"
• "Reason through step by step..."

Critical Triggers:

• "What could go wrong with this approach?"
• "Steelman the opposing view..."
• "What's the strongest argument against?"
• "What assumptions am I making?"
• "Where might this reasoning break down?"

Synthesis Triggers:

• "How do these pieces fit together?"
• "What's the underlying pattern?"
• "Synthesize into a coherent view..."
• "What's the essence of this?"
• "How would you explain this to an expert?"

Output Quality Indicators

Signs extended thinking is working well:

• Response acknowledges complexity and tradeoffs
• Multiple perspectives are genuinely considered
• Reasoning is visible, not just conclusions
• Uncertainty is explicitly addressed
• Edge cases and risks are surfaced proactively
• Recommendations are nuanced, not absolute

Signs it might need more prompting:

• Quick, surface-level response
• Only one perspective considered
• Confident without showing reasoning
• Missing obvious complexities
• Generic advice without context fit

Combining with Other Patterns

Extended Thinking + Agentic Execution

"First, think through the best approach to [task].
Consider alternatives and tradeoffs.
Once you have a clear plan, execute it autonomously,
but note any decision points where you chose between options."

Extended Thinking + Iterative Refinement

"Draft an initial approach to [problem].
Then critique your own draft - what could be better?
Revise based on your critique.
Show me both the final version and key improvements you made."

Extended Thinking + User Collaboration

"Analyze [situation] and identify the key decision points.
For each decision, explain the tradeoffs.
Before proceeding, I'll tell you my priorities,
then you can recommend the path that fits best."

When NOT to Use Extended Thinking

Some tasks benefit from quick, direct responses:

• Simple factual questions
• Straightforward code snippets
• Clear instructions with no ambiguity
• Tasks where speed matters more than depth
• Follow-up execution after thinking is done

For these, don't add complexity. Just ask directly.

Model notes (current family)

• Fable 5 — Anthropic's most capable model for the hardest reasoning and long-horizon agentic work; thinking is always on. Reserve for genuinely demanding tasks (it is priced above Opus tier).
• Opus 4.8 — the default for strong reasoning and agentic work; adaptive thinking + effort.
• Sonnet 4.6 — best balance of speed and intelligence; supports adaptive thinking.
• Haiku 4.5 — fastest/cheapest for simple, latency-sensitive tasks (no xhigh/max effort).

All current models excel at sustained reasoning chains, self-correction when given space, nuanced "it depends" answers, and multi-factor synthesis — raise effort before adding prose scaffolding when you need more rigor.

Integration with FrankX System

Use extended thinking for:

• Starlight Council decisions - Strategic synthesis across domains
• Architecture planning - System design for Arcanea, FrankX projects
• Content strategy - Weighing positioning, audience, platform fit
• Creative development - Book outlines, course structures, music direction
• Technical decisions - Framework choices, MCP design, agent orchestration

Extended thinking is about quality of cognition, not length of output. The best responses come from genuine deliberation - use these patterns to unlock that capability.