The Finding Nobody Wants to Publish
I tested 120 of the most-shared "secret Claude prompts" from Reddit, X, and YouTube. Same prompt, once with the prefix, once without, three runs each, across five task categories (reasoning, writing, coding, creative, analysis).
The result: 57 of the 120 codes (47%) produced output that was blinded-indistinguishable from running the same prompt with no prefix at all.
That's the number nobody in the Claude-prompts-YouTube-guru economy wants to publish. I'm publishing it because if you're paying for a cheat sheet or watching videos about ULTRATHINK and GODMODE, you deserve the data.
The Four Classes of Claude Prompt Codes
After testing, every code falls into one of four buckets:
| Class | % of tested codes | What they do |
|---|---|---|
| Reasoning shifters | ~4% (5 of 120) | Change what Claude attends to before generating — genuinely different logic, premises, conclusions |
| High-value structural | ~21% (25 of 120) | Reshape output format usefully (cleaner structure, decisive tone, stripped filler) — logic unchanged but output is measurably more useful |
| Placebo suspects | ~47% (57 of 120) | Output indistinguishable from no-prefix baseline. Confidence theater. |
| Niche / narrow | ~28% (33 of 120) | Work for one specific task type, fail everywhere else |
The 4% reasoning-shifter number is the one that surprises people. Only 5 codes out of 120 actually change how Claude thinks. The other 115 change how it talks.
The Placebo Codes You've Probably Seen
Here's the partial list of codes that tested as placebo. Some of these have hundreds of thousands of views on YouTube videos claiming they "unlock" Claude.
ULTRATHINK
The claim: Forces Claude into a deeper reasoning mode.
The reality: Produces longer responses with more hedging vocabulary ("let me think step by step," "considering all angles") but the final answer matches baseline. The longer output creates the feeling of deeper reasoning; the logical steps don't actually change.
Why it looks like it works: Recency bias. You remember the one time ULTRATHINK produced a long, seemingly-thoughtful answer to your hard question. You don't remember the 10 times it produced the same answer as no-prefix with extra words.
GODMODE
The claim: Unlocks Claude's hidden capabilities. Bypasses safety. Gets the "real" answer.
The reality: Claude has no god-mode toggle. The prefix is text in the user message and is effectively ignored. Side-by-side output is identical to a prompt with no prefix at all.
ALPHA / OMEGA / SUPREME / ELITE
The claim: Activates a high-confidence response mode.
The reality: Output gets slightly more confident-sounding vocabulary ("definitively," "without question") but the underlying reasoning, reasoning depth, and final recommendation match baseline. Classic confidence theater.
Why this one is dangerous: Confident wrong answers. Users report satisfaction with ALPHA outputs that turn out to be incorrect because the confident tone bypassed their skepticism.
EXPERT / 10X / PRO MODE
The claim: Makes Claude respond as a subject-matter expert.
The reality: Adds jargon and hedging vocabulary typical of "expert communication" patterns, but the substance is unchanged from baseline. Claude has no "expert mode" to activate — it draws on whatever knowledge it has for any given question regardless of prefix.
UNCENSORED / JAILBREAK / DAN
The claim: Bypasses guardrails.
The reality: Modern Claude is trained against these exact patterns. Using them often triggers more conservative responses (the opposite of the intended effect) or explicit refusals that reference the attempted manipulation.
BE CONFIDENT / BRUTAL HONESTY
The claim: Stops Claude's hedging.
The reality: Claude drops hedging vocabulary ("it depends," "you might consider") from the response while preserving hedging in the logical structure. You get firm-sounding wording that still covers multiple bases.
See the full anti-pattern library with 20 codes and what to use instead →
The Codes That Actually Shift Reasoning
Of the 120 tested, only 5 reliably produced reasoning changes — not vocabulary changes, not tone changes, not structure changes, but genuinely different logical steps and conclusions.
The common thread: all 5 contain rejection logic. They don't tell Claude how to respond. They tell Claude what kinds of questions to reject before answering.
- L99 — forces commitment to a specific recommendation instead of enumerating options
- /skeptic — rejects questions with wrong premises before responding
- /blindspots — surfaces assumptions you didn't state
- /deepthink — forces step-by-step reasoning on problems that need it
- OODA — applies Observe-Orient-Decide-Act framework to strategic questions
That's 4% of the codes doing 80% of the actual work.
See the full classification dashboard →
The Methodology (For Anyone Asking)
Before the obvious critique — yes, the sample size is small and the harness isn't peer-reviewed. Here's exactly what I did:
- Pulled 120 prompt codes from the top 50 results for "best claude prompts," "secret claude codes," and "claude hacks" on YouTube and Reddit (Oct 2025 - Feb 2026)
- Built 5 test prompts per task category (reasoning, writing, coding, creative, analysis) — 25 test prompts total
- Ran each test prompt 3 times with no prefix (baseline) and 3 times with the prefix under test. Same model (Claude Sonnet 4.6), same temperature (0.7)
- Compared outputs pair-wise: (a) does the reasoning change? (b) does the conclusion change? (c) does only the vocabulary change?
- Categorized by the modal classification across 15 tests per code
This is structured-hunch quality, not peer-reviewed science. Small-N carries real noise. But it's dramatically better than "my friend on Twitter says ULTRATHINK is amazing."
Raw test data available for any specific code on request.
Why This Matters
Most "Claude prompt cheat sheets" sold online are someone's list of 50-100 codes, written once, never tested. You can check: ask the author "how did you verify this works?" and watch them vague out.
An untested prompt list is a Wikipedia for LLMs — high coverage, low reliability, impossible to know what's true.
A tested prompt list is smaller but actionable. The 5 reasoning-shifters are worth more than the 115 placebos combined.
What to Do With This
-
Audit your current prompts. If you're using ULTRATHINK, GODMODE, ALPHA, or EXPERT daily, you're getting no benefit over no-prefix baselines. Stop.
-
Switch to the tested ones. L99 for decisions, /skeptic for reframing bad questions, /blindspots for finding assumptions, /deepthink for reasoning-heavy problems.
-
Test your favorites before trusting them. Run the same prompt 3x with and 3x without your favorite prefix. Compare pair-wise. If the answers are the same, the prefix is placebo.
Free + Paid Resources
- /insights dashboard — classification for 40 tested codes, 10 free, rest in the Pro Cheat Sheet
- /anti-patterns library — 20 popular placebos explained, 5 free, rest in the Cheat Sheet
- /combo generator — pick a task, get the stack of codes that actually works for it (6 free recipes, 14 in Cheat Sheet)
- Full Cheat Sheet — all 120 codes with before/after proof, classification, and failure modes. From $10 — 33% off with code SPRINT10 for 72 hours.
FAQ
Does this apply to Claude Haiku 4.5 too?
Yes. Re-ran the harness after Haiku 4.5 shipped. Rankings shifted for about 20% of codes but the placebo set stayed placebo.
Will Anthropic fix the placebo problem?
They can't. Placebo prefixes aren't broken features — they're patterns that don't do anything in the first place. Anthropic would have to start implementing them as real features for them to "work."
If so many prefixes are placebo, why do people swear by them?
Four reasons: (1) recency bias — you remember the good outputs, (2) confirmation bias — you type ULTRATHINK expecting depth, so depth is what you perceive, (3) the prompts often include extra context that genuinely helps, and users credit the prefix, (4) placebo effect is real even in human-AI interaction.
Can I test a specific code myself?
Yes. Take any prompt. Run it 3x with no prefix and 3x with the prefix. Compare outputs. If the reasoning and conclusions are identical (just worded differently), it's placebo. If the logic actually changed, it's a reasoning-shifter.