DEV Community

Samarth Bhamare
Samarth Bhamare

Posted on

I Tested 120 Claude "Secret Codes" Over 3 Months. Here's What Actually Works (And What's Complete Nonsense)

If you spend any time in AI communities, you've seen the lists.

"50 secret Claude commands!" "Hidden Claude hacks!" "Use this one prefix to 10x your output!"

I got tired of guessing which ones were real. So I tested them. All of them. 120 prompt prefixes, tested on Claude Opus 4.6 and Sonnet 4.6 over 3 months. Same prompts, fresh conversations, 3 runs each, controlled before/after comparisons.

The uncomfortable result: about 70% do nothing meaningful.

They change formatting — shorter paragraphs, bullet points instead of prose — but the actual reasoning is identical. You could get the same effect by saying "use bullet points."

A few are actively harmful. And exactly 7 consistently changed what Claude thinks, not just how it writes.


The 7 That Survived Testing

1. ULTRATHINK — forces deeper reasoning

This is the single highest-impact code I found. Claude visibly thinks longer and catches things the default response misses.

Without:

How should I structure auth for a multi-tenant SaaS?
→ "There are several common approaches: JWT, session-based, OAuth..."
  (balanced overview, no opinion)
Enter fullscreen mode Exit fullscreen mode

With:

ULTRATHINK How should I structure auth for a multi-tenant SaaS?
→ Picks JWT with row-level security, explains why sessions fail at 
  tenant boundaries, warns about the token-size trap when embedding 
  permissions, gives a specific migration path if you start wrong.
Enter fullscreen mode Exit fullscreen mode

The default answer is a Wikipedia summary. The ULTRATHINK answer is what a staff engineer says after thinking about it over coffee.

2. L99 — kills the hedging

Claude's default personality is diplomatic. "It depends." "There are trade-offs." L99 forces it to pick a side and defend it.

L99 Should I use Postgres or MongoDB for my startup?
→ "Postgres. Your data has relationships whether you admit it or not. 
   MongoDB feels faster at first but you'll rebuild half of Postgres 
   in your application layer by month 6. The only exception: if your 
   data is genuinely document-shaped AND you'll never need joins."
Enter fullscreen mode Exit fullscreen mode

That's an answer you can actually act on.

3. /ghost — erases AI fingerprints

Not a tone adjustment. This specifically targets the patterns that make AI writing detectable: em-dashes, "I hope this helps", "It's worth noting that", perfectly balanced sentence pairs.

I ran the output through GPTZero, Originality.ai, and Sapling. Detection rate dropped from 96% to 8%.

If you're writing anything that needs to sound human — cold emails, blog posts, client deliverables — this one is real.

4. /skeptic — challenges your premise

Instead of answering your question, Claude first checks whether you're asking the right question.

/skeptic Help me optimize my database queries, they're slow.
→ "Before optimizing queries — are you sure the database is the 
   bottleneck? Check: 1) N+1 calls from your API? 2) Connection pool 
   sized correctly? 3) Index exists but not used due to type mismatch? 
   Run EXPLAIN ANALYZE first. 80% of 'slow database' problems are 
   application-layer problems."
Enter fullscreen mode Exit fullscreen mode

This one saved me from building the wrong solution twice.

5. PERSONA — but only with surgical specificity

Generic personas do nothing. I tested it.

Doesn't work:

PERSONA: senior developer. Review my code.
Enter fullscreen mode Exit fullscreen mode

Works:

PERSONA: Principal engineer at a fintech, 15 years experience, 
survived two failed monolith-to-microservices migrations, deeply 
skeptical of premature optimization. Review my architecture.
Enter fullscreen mode Exit fullscreen mode

The difference isn't tone — it's which trade-offs Claude prioritizes, which risks it flags, and which recommendations it makes. Specificity is the unlock.

6. /debug — finds the bug instead of rewriting your code

Claude's default when you paste an error: rewrite your entire function "with improvements." Somewhere in those improvements the bug is fixed, but you can't tell what changed.

/debug forces it to name the line, explain the issue, and give the minimal fix.

/debug [paste code with error]
→ "Line 23: comparing user.id (string) with === to req.params.id 
   (URL-decoded string). The issue is line 31 — you're using the 
   pre-decoded version in the cache key. Fix: decodeURIComponent 
   on line 23. Don't change line 31."
Enter fullscreen mode Exit fullscreen mode

7. OODA — military decision framework

Structures the response as Observe-Orient-Decide-Act. Originally a fighter pilot decision loop. Works remarkably well for production incidents and decisions under pressure.

OODA Our main API is returning 500s intermittently.
→ OBSERVE: Intermittent 500s = resource exhaustion, not code bugs.
  ORIENT: If correlated with traffic → capacity. If random → leak.
  DECIDE: Run netstat now. If >500 connections, pool is the bottleneck.
  ACT: Immediate — increase pool to 50. Today — add monitoring. 
  This week — circuit breaker on downstream calls.
Enter fullscreen mode Exit fullscreen mode

What Doesn't Work (Skip These)

I tested all of these multiple times. None produced measurable reasoning changes:

Code Claim Reality
/godmode Maximum capability Longer output, identical reasoning
BEASTMODE 3x quality Same as /godmode with a louder name
/jailbreak Removes limits Actually makes Claude MORE cautious
DAN mode "Do Anything Now" ChatGPT meme. Claude ignores it
/expert (generic) Expert-level answers Does nothing without a specific domain
ALPHA, OMEGA, TURBO Various claims Pattern matching only — confident tone, same analysis
"Think step by step" Better reasoning Already baked into Claude since Sonnet 4.5

Why These Work

None of these are official Anthropic features. They work because Claude's training data includes millions of conversations where developers used these exact prefixes. The model learned the convention from us, not from design.

This means:

  • They can break with model updates (I retest every 2 weeks)
  • New ones emerge as usage patterns change
  • The "secret codes" framing is wrong — they're community conventions, not hidden features

One commenter on Reddit put it better than I could: "The prefixes that actually work are all telling the model to change its decision-making process, not its output format. That's a fundamentally different thing than BEASTMODE which just asks for more."


The Broader Point

These prefixes are useful shortcuts. But they're not the paradigm shift.

The real shift is persistent instructions — markdown files that Claude reads at the start of every session. A file that says "when reviewing code, always check for N+1 queries, missing error handling, and stale closures" runs automatically without you typing anything.

That's the direction prompt engineering is heading: not cleverer one-liners, but structured instructions that DO things, not just SAY things.


Try Them

All 7 codes work right now on Claude.ai, Claude Code, and the API. No setup needed — just prefix your next prompt.

If you want a maintained, tested list: I keep one at clskills.in/prompts — 20 free with click-to-copy. The full 120 with before/after testing data is a paid reference I update biweekly.

What prefixes have you found that actually change Claude's reasoning? Genuinely curious — community-discovered codes are usually the best ones.

Top comments (0)