ULTRATHINK Claude Prompt — Maximum Reasoning Depth, Zero Hedging
ULTRATHINK is the most powerful reasoning prefix for Claude — deeper than L99, longer answers, zero hedging. Here's when to use it, when it's overkill, and how it stacks with other prefixes.
Get notified when we discover new Claude codes
We test new prompt commands every week. Join 4+ developers getting them in their inbox.
What ULTRATHINK actually does
ULTRATHINK is the nuclear option for Claude reasoning depth. If L99 pushes Claude to commit to an opinion, ULTRATHINK pushes Claude to build an entire thesis — 4-8 paragraphs of deep, layered analysis with zero hedging and zero "it depends."
The output feels like a senior consultant who's been thinking about your problem for an hour, not 10 seconds. You'll get the surface answer, then 3-4 layers underneath, then the structural cause, then what to do about it.
How it compares to other reasoning prefixes
The Claude community has discovered several reasoning-depth prefixes. Here's how they compare, ranked from lightest to deepest:
| Prefix | Depth | Best for |
|---|---|---|
/deepthink | Medium | Showing the work — walks through each reasoning step |
L99 | High | Getting a committed opinion instead of "it depends" |
THINK_HARDER | Higher | Double-pass — generates an answer, critiques it, revises |
ULTRATHINK | Maximum | Full thesis with multiple layers of analysis |
THINK_HARDEST | Maximum (3-pass) | Triple-pass — three rounds of self-critique. Slowest but most thorough |
The key difference: L99 changes the tone (more committed, less hedging). ULTRATHINK changes the structure (longer, deeper, more layers). They stack well together, but for most purposes ULTRATHINK alone already implies the commitments L99 gives.
Before/after example
Question: Should I rewrite my Express.js backend in Go?
Default Claude: A balanced 4-paragraph essay listing pros and cons, ending with "ultimately it depends on your team's familiarity with Go and your performance requirements."
With L99: Commits to a recommendation: "Don't rewrite. Here's why, and here are the 3 conditions under which I'd change my mind." About 400 words.
With ULTRATHINK: Full thesis in 800-1200 words:
Surface answer: Don't rewrite. Express handles your current scale fine.
Layer 1 — the performance question: Go is ~3-5x faster at CPU-bound workloads, but your bottleneck is almost certainly I/O (database queries, API calls), not CPU. Node.js handles I/O-bound workloads just as well as Go. Rewriting for performance only makes sense if you've profiled and confirmed you're CPU-bound.
Layer 2 — the real cost: The rewrite itself will take 4-6 months for a team of 2. During that time, feature velocity drops to near-zero. Your competitors ship 2 quarters of features while you're rewriting something that already works. The business cost of the rewrite is measured in lost market opportunity, not engineering hours.
Layer 3 — the hidden assumption: You're assuming the backend is the problem. Have you checked? Most "my backend is slow" problems are actually "my database queries are unoptimized" or "my frontend is making too many requests" problems. A rewrite won't fix those — they'll just be the same problems in a new language.
Layer 4 — the structural cause: The urge to rewrite usually comes from one of three places: (1) a new CTO who prefers Go, (2) hiring pressure ("Go engineers are easier to find"), or (3) resume-driven development ("I want Go on my resume"). All three are valid motivations but none of them are technical reasons. Name which one is driving this and you'll know whether the rewrite is a business decision or a technical one.
Recommendation: Don't rewrite. Instead, profile the actual bottlenecks, optimize the top 3, and revisit in 6 months. If you're still CPU-bound after optimization, then the rewrite conversation is real.
That's the depth difference. Default Claude gives you the textbook answer. L99 gives you a committed opinion. ULTRATHINK gives you the full analysis a senior engineer would produce after a week of thinking.
When to use ULTRATHINK
ULTRATHINK is the right tool when the question earns the depth:
1. High-stakes technical decisions
"Should we move to microservices?" "Should we switch databases?" "Should we adopt this new framework?" These are decisions that affect the next 1-2 years of your team's work. They earn the depth.
2. Strategy and business decisions
"Should we raise funding?" "Should we pivot?" "Should we hire a sales team?" ULTRATHINK produces the kind of memo you'd want before making a major commitment.
3. Root cause analysis
"Why is our churn rate 8%?" "Why are deploys failing every Friday?" "Why can't we hire senior engineers?" ULTRATHINK goes past the obvious answer to the structural cause underneath.
4. When you want a second opinion on your own conclusion
You've already decided what you think. Run ULTRATHINK on the same question and see if it reaches the same conclusion through a different path. If it does, you're probably right. If it doesn't, the divergence tells you something.
When NOT to use ULTRATHINK
Don't ULTRATHINK simple questions
"How do I center a div?" "What's the syntax for a Python dictionary?" The depth is wasted and the latency (30-60 seconds) is annoying.
Don't ULTRATHINK urgent questions
If you're in an incident and need a quick answer, ULTRATHINK's thoroughness is friction. Use OODA for incidents — it gives you a runbook in 10 seconds.
Don't ULTRATHINK everything
The most common ULTRATHINK mistake is using it as your default. It produces 3-5x more text than a normal response. If you ULTRATHINK every question in a day, you'll spend more time reading answers than doing work. Reserve it for the 2-3 decisions per week that actually matter.
Don't ULTRATHINK when you want multiple perspectives
ULTRATHINK commits hard to one thesis. If you want to explore different angles, use /layered (surface/mid/expert levels) or IDEATE (20 ideas before evaluating) instead.
ULTRATHINK combos that stack
ULTRATHINK + /blindspots
The deepest analysis you can get. ULTRATHINK builds the thesis; /blindspots then asks "what did this thesis assume that might not be true?" Together they produce a recommendation AND its failure modes.
ULTRATHINK + PERSONA
Load a specific expert persona, then ask them to ULTRATHINK a question. "PERSONA: CTO who just went through a failed microservices migration. ULTRATHINK: should we move to microservices?" The persona constrains the analysis to a specific experience base, which prevents the answer from being generic.
ULTRATHINK + /skeptic
/skeptic challenges whether you're asking the right question. ULTRATHINK builds the deepest possible answer IF the question survives the challenge. Together they ensure you're solving the right problem at maximum depth.
ULTRATHINK + WORSTCASE
ULTRATHINK builds the recommendation. WORSTCASE then asks "what's the worst that happens if we follow this recommendation and it's wrong?" Best for irreversible decisions where the cost of being wrong is high.
ULTRATHINK vs /godmode
People often confuse these. The difference:
- ULTRATHINK = maximum reasoning depth. The answer has layers, each layer goes deeper into the structural cause.
- /godmode = maximum comprehensiveness. The answer covers everything — history, mechanics, alternatives, resources, edge cases. More like a textbook chapter than a thesis.
Use ULTRATHINK for decisions. Use /godmode for learning a topic. Use both together if you want 3,000 words of the most thorough analysis Claude can produce (but be prepared to spend 5 minutes reading it).
Where to find more
ULTRATHINK is one of about 120 Claude prompt prefixes the community has discovered. The first 11 are free at clskills.in/prompts.
The full set with before/after examples for every code, when-NOT-to-use warnings, and the combos that stack: Claude Code Cheat Sheet — one-time $5-$10 purchase, lifetime updates.
Related deep dives:
- L99 Claude Explained
- PERSONA Claude Prompt
- /skeptic Claude Prompt
- /ghost Claude Prompt
- Claude Mythos Prompt Guide — what to expect when the next model drops
TL;DR
ULTRATHINK is the maximum-depth reasoning prefix. Deeper than L99, produces 800-1200 word thesis-style responses with 3-4 layers of analysis. Zero hedging, zero "it depends."
Use it for: high-stakes decisions, root cause analysis, strategy memos, getting a second opinion on your own conclusion.
Skip it for: simple questions, urgent situations, anything where reading 1000 words is more expensive than getting a slightly shallower answer.