DEV Community

Samarth Bhamare
Samarth Bhamare

Posted on • Originally published at clskills.in

Claude "Cheat Codes" — 120 Tested Prompt Prefixes That Change How Claude Responds

You can ask Claude the same question two different ways and get two completely different answers.

I don't mean rephrasing. I mean adding one tiny prefix in front of your prompt. The community has converged on about 120 of these "cheat codes" — none of them are official Anthropic features, but Claude consistently recognizes them.

I spent 3 months testing which ones actually work. Here's the complete breakdown.

What are Claude cheat codes?

Short commands you put at the start of your message:

  • Slash commands/ghost, /skeptic, /punch, /noyap
  • Uppercase keywordsL99, ULTRATHINK, OODA, PERSONA, BEASTMODE

Claude sees the prefix, recognizes the intent, and adjusts its behavior.

The 5 you should learn first

L99

Forces committed opinions instead of "it depends."

L99 Should I use Postgres or MongoDB for a fintech startup?
Enter fullscreen mode Exit fullscreen mode

You get a real recommendation with specific tradeoffs, not a balanced pros-and-cons essay.

PERSONA (with specificity)

Generic: "Act like a database expert" → barely moves the needle.

Specific:

PERSONA: Senior Postgres DBA at Stripe, 15 years, 
cynical about ORMs. How do I optimize this query?
Enter fullscreen mode Exit fullscreen mode

The specificity gives Claude permission to have real opinions.

/ghost

Strips every "AI tell" from writing. Em-dashes, "I hope this helps", balanced sentence pairs — all gone. Output reads like a human typed it.

/ghost Rewrite this so it doesn't sound like AI.
Enter fullscreen mode Exit fullscreen mode

/skeptic

Challenges your question before answering it.

/skeptic How do I A/B test 200 landing page variants?
Enter fullscreen mode Exit fullscreen mode

Claude responds: "With your traffic, you'd need 6 months per variant. Want to test 5 instead?"

Catches the wrong-question problem.

ULTRATHINK

Maximum reasoning depth. 800–1200 word thesis with multiple analytical layers. Zero hedging.

ULTRATHINK Why is our retention dropping after week 2?
Enter fullscreen mode Exit fullscreen mode

Reserve for the 2–3 decisions per week that earn the depth.

Codes that stack (combos)

Most prefixes get stronger combined:

Combo Best for
/ghost + /punch Cold emails — humanize + shorten
PERSONA + L99 Technical decisions — expert + committed
/skeptic + ULTRATHINK Strategy — right question + max depth
/punch + /trim + /raw Slack messages — tight, no markdown

The full 120

I organized all of them into a searchable reference with categories:

Category Count Examples
Writing & Style 10 /ghost, /punch, /trim, /voice
Thinking & Reasoning 10 L99, /deepthink, OODA, XRAY
Coding & Technical 10 /debug, REFACTOR, /shipit, ARCHITECT
Power Commands 10 /godmode, PERSONA, /memory, /chain
Hidden Modes 10 BEASTMODE, /nofilter, CEOMODE, OPERATOR
Output Formatting 10 /json, /yaml, /table, /checklist
Research & Analysis 10 DEEPDIVE, COMPARE, /critique, /devil
Productivity 10 /template, /recipe, /playbook, /sop
Communication 10 /email, /reply, /dm, /escalate
Experimental 10 PARETO, FINISH, /postmortem, WORSTCASE
Community Picks 20 ULTRATHINK, /noyap, HARDMODE, /skeptic

Free interactive version: clskills.in/prompts — 11 free codes, click to copy.

Individual deep dives: clskills.in/prompt/l99, /prompt/ghost, /prompt/persona, /prompt/skeptic — 120 pages, one per code.

Full cheat sheet (paid, $5–$10): clskills.in/cheat-sheet — before/after examples for every code, when-NOT-to-use warnings, combos, 10 workflow playbooks. Lifetime updates.

Do these actually work or is it placebo?

Honest answer: these are community conventions, not hardcoded features. They work because Claude has seen enough people use them that it learned to recognize the intent.

Evidence they work:

  1. Output structure measurably changes (run the same prompt with/without L99 in fresh sessions)
  2. Multiple people discovered the same codes independently
  3. Anthropic hasn't denied them

Caveats:

  1. Top 15 are very consistent. Bottom 30 are more hit-or-miss.
  2. Future model updates could change behavior.
  3. Some codes work better for certain task types.

Try it now

Pick one:

  • L99 [any question you'd normally ask Claude]
  • /ghost [any paragraph you want rewritten]
  • PERSONA: [specific expert]. [your question]

If it doesn't change anything, the rest won't either. If it does — the other 117 are waiting.


Building clskills.in — 2,300+ free Claude Code skills + the cheat sheet for prompt prefixes.

Top comments (0)