7 Hidden Claude Prompts That Actually Work (Tested in 2026)

Most "Secret Claude Prompts" Are Fake

If you've spent any time in AI communities, you've seen the lists. "50 secret Claude prompts!" "Claude jailbreak codes!" "Use this hidden command to unlock GPT-4 in Claude!"

Most of them don't work. We tested 50+ supposedly-secret prompts over the last 6 weeks. Most do nothing. A few do the opposite of what they claim. And a small number — maybe 7 — genuinely change Claude's behavior in useful ways.

This is that list. Every code below has a real before/after example we ran ourselves.

1. L99 — The Depth Code

Type L99 at the start of a prompt. Claude switches into a noticeably deeper, less hedged response style.

Before: "How should I structure a Postgres database for a multi-tenant SaaS?" → Claude gives a 200-word generic overview of three options.

After: L99 How should I structure a Postgres database for a multi-tenant SaaS? → Claude gives an 800-word opinion-first answer, explains exactly why one approach wins for your stage, walks through the migration cost of getting it wrong, and ends with a specific schema example.

Full L99 explainer →

2. /ghost — Removes the AI Tells

Type /ghost and Claude strips out all the meta-commentary that makes AI writing feel like AI writing. No "Here's what I think," no "It's worth noting," no "In summary."

Before: "Write a tagline for my startup." → "Here are several tagline options for your consideration: 1. ... 2. ... I hope these help! Let me know if you'd like variations."

After: /ghost Write a tagline for my startup. → "Ship faster. Sleep better."

The difference is night and day if you're using Claude for any kind of creative writing.

3. /deepthink — The Reasoning Trigger

/deepthink forces Claude to reason through every layer of a problem before answering. Slower output, dramatically better quality on complex questions.

Before: "Why is my React component re-rendering on every state change?" → Claude lists the 5 most common causes.

After: /deepthink Why is my React component re-rendering on every state change? → Claude actually walks through your code, hypothesizes which cause is most likely based on patterns it sees, and explains its reasoning before giving the answer.

Use this for debugging, architecture, and any "why" question where the obvious answer is probably wrong.

4. OODA — The Decision Framework

Start your prompt with OODA and Claude structures the response as Observe-Orient-Decide-Act. Originally a military decision loop. Works incredibly well for tech decisions under uncertainty.

Use it when: you need to decide something and you don't have time for an essay.

Best for: production incidents, architecture decisions, postmortems.

Full OODA breakdown →

5. ARTIFACTS — The Deliverables Mode

ARTIFACTS tells Claude to structure output as a numbered list of concrete deliverables. Instead of explaining what you should build, Claude lists exactly what it'll produce, then produces each one.

Before: "Help me launch a landing page." → Claude explains landing page best practices for 500 words.

After: ARTIFACTS Help me launch a landing page. → "ARTIFACT 1: Hero copy. ARTIFACT 2: Feature grid. ARTIFACT 3: Pricing table. ARTIFACT 4: HTML implementation." Then it builds each one.

6. /mirror — The Style Lock

Give Claude a sample of your writing and prefix the next prompt with /mirror. It'll match your voice, sentence rhythm, and vocabulary instead of defaulting to its own style.

Use case: writing sales emails in your own voice, drafting tweets that sound like you, ghostwriting blog posts that don't sound AI-generated.

This one is genuinely useful for anyone who writes content.

7. PERSONA — The Expert Channel

PERSONA: <role> tells Claude to respond as a specific expert. The trick: you have to be specific. "PERSONA: senior developer" doesn't do much. "PERSONA: senior backend engineer who specialized in PostgreSQL performance for 15 years at high-scale fintech" gets you a different answer entirely.

Before: "How do I make my queries faster?" → Generic checklist.

After: PERSONA: Postgres DBA with 15 years at Stripe. How do I make my queries faster? → Specific advice about pg_stat_statements, prepared statement caching, connection pool sizing for your scale, and 3 things "most people get wrong" — all in the voice of someone who's actually done it.

What Doesn't Work (Skip These)

We also tested these and found they do nothing or actively hurt response quality:

• /jailbreak — Makes Claude MORE cautious, not less.
• DAN mode — This is a ChatGPT meme, Claude has no idea what you mean.
• /uncensored — Same as above.
• /godmode (with no other context) — Claude just outputs more words, not better ones.
• /expert (without specifying which kind) — Useless without a real persona.
• Random uppercase strings like ALPHA, OMEGA, MAX — Pattern-matching only, no real effect.

If someone tells you these are "hidden Claude commands," they haven't tested them.

The Pattern Behind Why These Work

None of the working codes are official. They work because Claude's training data includes thousands of developers using these exact patterns in real prompts. The model learned the convention from us, not from Anthropic.

This also means: as Claude updates, some codes will stop working and new ones will emerge. We test the full list every two weeks and update clskills.in/prompts with what's currently working.

Stay Updated

We find new working prompt codes every week — and mark old ones as deprecated when models update. Subscribe to the newsletter at the top or bottom of this page to get them as we test them.

Or browse all 100+ tested codes in our searchable cheat sheet.