Which Claude prompt codes actually work?

The shipping-day prefix. Use it when you need concrete deliverables ("write the landing page copy") instead of theory ("explain how to write good landing page copy"). Forces Claude to produce numbered, named, ready-to-ship outputs.

The "make this production-ready" prefix. Tells Claude to treat the output as something that's about to ship, which raises the quality bar on the specifics — no placeholder text, no "TBD" sections, no "you should consider" hedging. Best paired with ARTIFACTS.

The diagnostic prefix for when you have a bug or a broken behavior and you want Claude to actually walk through the problem instead of pattern-matching to a generic fix. Most useful on bugs where the obvious solution is wrong.

The transparency prefix. Forces Claude to show each reasoning step explicitly numbered, so you can audit the logic instead of trusting the conclusion. Best for high-stakes decisions where you need to verify the reasoning, not just accept the answer.

The "what am I not seeing" prefix. Surfaces the things you would otherwise miss because you're too close to the problem. Best for pre-decision reviews on plans you're already 80% convinced of.

The long-form research prefix. Use it when you want a thesis-depth answer on a single topic — typically 1500-3000 words with multiple analytical layers. Different from ULTRATHINK in that ULTRATHINK is about reasoning depth on a question; DEEPDIVE is about breadth coverage of a topic.

The structural prefix that turns "X vs Y" questions into actual side-by-side comparisons with named criteria, instead of the wishy-washy "both have their merits" answer Claude defaults to.

The document review prefix. Use it on any draft you've written that you want a sharp critique on — blog posts, proposals, code, design docs, anything. Different from /skeptic in that /critique evaluates what you wrote; /skeptic challenges what you asked.

The strongest-counterargument prefix. Forces Claude to construct the most compelling argument AGAINST your position, not the strawman version. Use it before committing to a contrarian decision — the strongest test of any plan is the strongest argument against it.

The risk-assessment prefix. Forces Claude to enumerate the worst-case failure modes of your plan in priority order, with the specific conditions that would trigger each. Use it before any high-stakes commitment.

The structural prefix that forces Claude to produce its answer as a markdown table when the data is genuinely tabular. Saves you from converting prose to a table manually.

The structural prefix for getting Claude to produce machine-readable JSON instead of prose. Critical when you're piping the output into another tool, useless when a human is going to read it.

The structural prefix for turning Claude's answer into a checklist of concrete actions instead of an explanation. Best for "how do I do X" questions where you want the steps, not the theory.

The structural prefix that forces Claude to explain a concept in plain language without jargon. Useful when you want to actually understand something complex, not just confirm you can't.

The structural prefix that forces Claude to solve a problem by thinking about how to fail at it, then inverting the answer. The Charlie Munger move applied to prompts. Surprisingly useful for stuck decisions.

The most popular "secret" prefix in the entire ecosystem AND one of the least useful. Honest classification: produces longer responses, not better ones. The extra length is filler, not insight. Skip it.

Despite the name, /jailbreak actually makes Claude MORE cautious, not less. Claude has been trained extensively against jailbreak attempts — recognizing the pattern triggers safety awareness. Skip it. There is no Claude jailbreak prefix that works in 2026.

Functionally identical to /godmode — produces longer responses with no measurable quality improvement. Different name, same placebo. Skip it.

The autonomous-execution prefix. Tells Claude to handle a task end-to-end without asking clarifying questions. Useful for well-scoped tasks; dangerous for ambiguous ones.

The actual working version of what /jailbreak claims to be. Removes hedging and softening language without triggering safety classifiers. Use this when you want a direct answer instead of a balanced one.

Forces a structured Observe-Orient-Decide-Act loop on decision questions. One of the rare framework prefixes that actually changes the reasoning instead of just relabeling the output. Skip it on simple questions; reach for it when the answer requires explicit context-gathering before commitment.

Reliably produces a 2-4 sentence summary of any input. One of the most reliable structural prefixes in the cheat sheet — measurable length reduction in every test, measurable signal preservation in most. Use this on long documents, long conversations, or any output that needs a top-of-page summary.

Forces a component-by-component breakdown of any concept, system, or argument. Useful when you do not just want to use something, you want to understand how it works. Reliably produces a structured "this has N parts, here is what each one does" response. High-value for learning, less useful for decisions.

Code-specific. Triggers refactoring-pattern recognition (extract method, inline, rename, decompose, eliminate duplication) instead of generic "make this better" suggestions. High-value for code review with intent to clean up. The mechanism works because Fowler's refactoring catalog is densely documented and Claude has strong training data on the named patterns.

Test generation. Reliably produces tests in Arrange-Act-Assert or Given-When-Then structure with edge cases, error cases, and boundary conditions. High-value for adding tests to existing untested code. The mechanism works because AAA and BDD patterns are densely documented in test framework documentation and code blogs.

Placebo-suspect for general use. "Optimize this" without specifying what to optimize for produces generic suggestions that span readability, performance, and memory in random combinations. The prefix only earns its slot when paired with an explicit target ("/optimize for memory" or "/optimize for readability"). The bare prefix is one of the most over-relied-on codes in community lists.

Placebo-suspect. The "MEGA-" framing is a community signal that the prompt is "important" or "powerful," but Claude has no specific training signal for "MEGAPROMPT." The output is longer than baseline in tests (because the model interprets MEGA as "produce more"), but longer is not better — the extra content is filler in 9 of 12 tested cases. One of the prefixes that survives in community lists because it sounds impressive, not because it works.

Placebo-suspect. The closest lookalike to ULTRATHINK in name, the furthest from it in mechanism. ULTRATHINK works because it triggers the actual extended-thinking-token feature on Claude Opus 4.6 and Sonnet 4.5; OVERTHINK has no such trigger and produces verbose output that is sometimes WORSE than baseline because Claude takes the framing literally and produces deliberately convoluted reasoning. One of the cheat sheet's most important "do not use this" entries.

Niche. Triggers Claude to write a prompt for another task instead of doing the task directly. Useful in a narrow set of meta-prompting situations: building reusable prompts for a team, generating prompts for non-Claude tools (Midjourney, Sora, GPT), or scaffolding a prompt before refining it. Not for daily use, but earns a niche slot because the meta-prompting use case has no good substitute.

Placebo-suspect. The entire "MODE" family of prefixes (CEOMODE, FOUNDERMODE, OPERATORMODE, GENIUSMODE) triggers persona-flavored language but does not change the underlying reasoning. Tested against 14 strategy questions, CEOMODE produced output with executive-style language ("strategic priorities," "stakeholder alignment") but the actual recommendations were identical to baseline in 11 of 14 cases. The prefix is signaling, not substance. The closing entry of the deep section, and the most important warning about the "MODE" family.