Claude Cheat Codes — The Complete Guide to Every Working Prompt Prefix (2026)
Every Claude "cheat code" that actually works in 2026. 120 community-discovered prompt prefixes tested over 3 months, organized by category, with examples. The definitive reference.
Get notified when we discover new Claude codes
We test new prompt commands every week. Join 4+ developers getting them in their inbox.
What are Claude cheat codes?
Claude "cheat codes" (also called prompt prefixes, secret codes, or prompt patterns) are short commands you type at the start of your message to change how Claude responds. They're not official Anthropic features — they're conventions the community has discovered through experimentation that Claude consistently recognizes and responds to.
There are two types:
- Slash commands —
/ghost,/skeptic,/punch,/noyap. Type with the forward slash. - Uppercase keywords —
L99,ULTRATHINK,OODA,PERSONA,BEASTMODE. Type in all caps.
Both work the same way: Claude sees the prefix, recognizes the intent, and adjusts its response accordingly.
How many cheat codes exist?
As of April 2026, the community has discovered about 120 working codes across 11 categories. Not all of them are equally useful — some are situational, some are powerful daily drivers, and some are experimental.
Here's the breakdown by category:
| Category | Codes | Examples |
|---|---|---|
| Writing & Style | 10 | /ghost, /punch, /trim, /voice, /hook |
| Thinking & Reasoning | 10 | L99, /deepthink, OODA, /blindspots, XRAY |
| Coding & Technical | 10 | /debug, REFACTOR, /shipit, ARCHITECT, /testit |
| Power Commands | 10 | /godmode, MEGAPROMPT, PERSONA, /memory, /chain |
| Hidden Modes | 10 | BEASTMODE, /nofilter, CEOMODE, OPERATOR, SENTINEL |
| Output Formatting | 10 | ARTIFACTS, /table, /json, /yaml, /checklist |
| Research & Analysis | 10 | DEEPDIVE, COMPARE, /critique, /devil, /steelman |
| Productivity Hacks | 10 | /template, /recipe, /cheatsheet, /playbook, /sop |
| Communication | 10 | /email, /reply, /dm, /followup, /escalate |
| Bonus / Experimental | 10 | PARETO, /eli5, FINISH, /postmortem, WORSTCASE |
| Community Picks (new) | 20 | ULTRATHINK, /noyap, HARDMODE, /skeptic, /devmode |
Total: 120 tested codes.
The 5 most important ones to learn first
If you only learn 5, make it these. They cover the most common use cases and produce the most dramatic quality improvement:
1. L99 — Stop the hedging
Default Claude hedges everything. "It depends on your needs." "There are several approaches." "Consider both options."
L99 before your question forces Claude to commit to one answer. You get a real recommendation, not a menu.
L99 Should I use Postgres or MongoDB for my fintech startup?
Result: A committed recommendation with the specific tradeoffs and the conditions that would change the answer.
2. /ghost — Remove AI fingerprints from writing
Claude's default writing has recognizable patterns: em-dashes everywhere, "I hope this helps", balanced sentence pairs, "In conclusion." Human readers have learned to spot these.
/ghost strips them. The output reads like a human wrote it.
/ghost Rewrite this paragraph so it doesn't sound like AI.
3. PERSONA — Expert-level answers (only with specific personas)
Generic personas ("act like an expert") barely change the output. Specific personas with biases and experience produce dramatically better answers.
PERSONA: Senior backend engineer at Stripe, 12 years, cynical about microservices. Review my architecture.
4. /skeptic — Challenge the question before answering
Most prompt codes improve the answer. /skeptic improves the question. It makes Claude challenge your premise before responding — catching the "wrong-question problem" before you waste time.
/skeptic Should I build a recommendation engine for my 500-user app?
5. ULTRATHINK — Maximum reasoning depth
The deepest thinking mode available. Produces 800-1200 word thesis-style responses with multiple analytical layers. Reserve for decisions that actually earn the depth.
ULTRATHINK Why is our user retention dropping after the first week?
How to actually use cheat codes
The format is always the same: prefix first, then your normal prompt.
[PREFIX] [your normal prompt]
Examples:
L99 What programming language should I learn next?/ghost Rewrite this email so it doesn't sound like AI.PERSONA: CTO of a 50-person startup. Review my technical roadmap./debug Here's my failing test: [paste test + code]OODA Production is down. 502 errors. Last deploy 3 hours ago.
Cheat codes that stack (combos)
Most codes can be combined. The community has discovered specific combos that produce results better than either code alone:
/ghost + /punch— humanize the writing AND make every sentence sharper. Best for cold emails.PERSONA + L99— expert perspective AND committed opinion. Best for technical decisions./skeptic + ULTRATHINK— challenge the question AND go maximum depth on the answer. Best for high-stakes strategy./punch + /trim + /raw— short, tight, no-markdown. Best for Slack messages.PERSONA + WORSTCASE— what would this specific expert be most worried about? Best for risk assessment.
Where to get all 120
Free (11 codes): Browse the first 11 at clskills.in/prompts — click to copy, no signup.
Full reference (120 codes): The Claude Code Cheat Sheet includes all 120 codes with before/after examples, when-NOT-to-use warnings, combos that stack, and 10 real-world workflow playbooks. Three tiers from $5 to $10, lifetime updates.
Do these codes actually work, or is it placebo?
Fair question. Here's the honest answer:
These are community-discovered conventions, not hardcoded features. Claude recognizes them because it's been trained on millions of conversations where people used these patterns. When Claude sees L99, it's learned to associate that prefix with "the user wants a deep, committed answer."
Some evidence they work:
- Anthropic hasn't denied them. They haven't officially documented them either, but they haven't said "these don't do anything."
- Output structure measurably changes. Run the same prompt with and without L99 in fresh conversations. The length, commitment level, and structure are consistently different.
- The community has converged independently. Multiple people discovered the same codes without coordinating. That's harder to explain if the codes have no effect.
Some honest caveats:
- Not all codes are equally reliable. The top 15 are very consistent. The bottom 30 are more hit-or-miss.
- Future model updates could change behavior. A code that works on Opus 4.6 might work differently on whatever comes next.
- Context matters. A code that works great for technical questions might be useless for creative writing.
The testing I did was specifically to separate the ones that consistently work from the ones that are placebo or inconsistent. The 120 in the cheat sheet are the survivors.
Related guides
- 100 Claude Secret Codes — The Full List
- L99 Claude Explained
- /ghost Claude Prompt Guide
- PERSONA Claude Prompt Guide
- /skeptic Claude Prompt Guide
- ULTRATHINK Claude Prompt Guide
- Claude Mythos Prompt Guide
- The 15 Best Claude Prompt Prefixes in 2026
TL;DR
Claude cheat codes are prompt prefixes that change how Claude responds. 120 have been tested and documented. Start with L99 (committed opinions), /ghost (human-sounding writing), PERSONA (expert answers), /skeptic (challenge the question), and ULTRATHINK (maximum reasoning depth). They're free to try at clskills.in/prompts, and the full deep-dive reference is at clskills.in/cheat-sheet.