The 15 Best Claude Prompt Prefixes in 2026 (Tested and Ranked)
The 15 Claude prompt prefixes that actually make a measurable difference in output quality. Ranked by usefulness, with examples for each. Tested over 3 months on Opus and Sonnet.
Get notified when we discover new Claude codes
We test new prompt commands every week. Join 4+ developers getting them in their inbox.
How we ranked these
Over the last 3 months, I tested 120 Claude prompt prefixes across hundreds of real conversations — technical decisions, writing tasks, code reviews, business strategy, customer emails. The methodology was simple: run the same prompt with and without each prefix in fresh conversations (no memory, no project context), and note which ones consistently changed the output in a measurable way.
These 15 are the ones that survived. They're ranked by how often I actually reach for them in daily use, not by how impressive they sound.
The top 15
1. L99
What it does: Forces Claude to commit to an opinion instead of hedging.
Why #1: This is the prefix I use most often. Default Claude answers "it depends on your needs" to almost every decision question. L99 stops that. You get a real recommendation, the specific tradeoffs, and the conditions under which Claude would change its mind.
Example: L99 Should I rewrite my Express backend in Go?
Deep dive: L99 Claude Explained
2. PERSONA
What it does: Makes Claude respond as a specific expert. Only works with absurdly specific personas — "Senior Postgres DBA at Stripe, 15 years, cynical about ORMs" beats "act like a database expert" by miles.
Why #2: The most dramatic output quality change of any prefix, IF you use it right. Generic personas ("act like an expert") barely move the needle. Specific personas with named biases and scars produce answers you'd pay $400/hour for.
Example: PERSONA: VP of Engineering who just survived a failed microservices migration. Should we adopt microservices?
Deep dive: PERSONA Claude Prompt Guide
3. /skeptic
What it does: Instead of answering your question, Claude challenges the question itself first. Catches the "wrong-question problem" — when you'd waste an hour acting on a correct answer to the wrong question.
Why #3: This is the prefix I wish I'd known about a year ago. I estimate it's saved me 20+ hours of work by catching bad assumptions before I acted on them.
Example: /skeptic How do I A/B test 200 variants of my headline?
Deep dive: /skeptic Claude Prompt Guide
4. /ghost
What it does: Strips every AI tell from the writing — em-dashes, "I hope this helps", balanced sentence rhythm, "In conclusion". Output reads like a tired human typed it.
Why #4: Essential for cold emails, Slack messages, social media, and any text where sounding like AI is actively harmful to your goals.
Example: /ghost [paste your AI-generated paragraph]
Deep dive: /ghost Claude Prompt Guide
5. ULTRATHINK
What it does: Maximum reasoning depth. Produces a 800-1200 word thesis-style response with 3-4 analytical layers. Zero hedging.
Why #5: The "I need Claude to actually think about this" button. Reserve for the 2-3 decisions per week that actually earn the depth.
Example: ULTRATHINK Why is our user retention dropping after week 2?
Deep dive: ULTRATHINK Claude Prompt Guide
6. OODA
What it does: Observe → Orient → Decide → Act. Best for time-pressured operational decisions where you need a runbook, not a lecture.
Why #6: The only prefix I use during actual incidents. When production is down and you're panicking, OODA gives you the 4-step runbook in 10 seconds.
Example: OODA Production API is returning 502s. Started 5 minutes ago. Last deploy was 2 hours ago.
7. /punch
What it does: Rewrites every sentence to be ~40% shorter and lead with a verb.
Why #7: The editing prefix I reach for most. "It might be worth considering whether we should perhaps look into refactoring the auth module" becomes "Refactor auth." Use it on anything you wrote when tired.
Example: /punch [paste your verbose paragraph]
8. /noyap
What it does: Pure answer mode. No preamble, no "great question", no "let me think about this". Just the answer in the first sentence.
Why #8: The best prefix for quick lookups. When you know what you want and don't need a tutorial.
Example: /noyap How do I sort a Python list of dicts by a key?
9. HARDMODE
What it does: Banishes "it depends" and "consider both options". Forces Claude to actually pick one and defend it.
Why #9: For decisions where you want commitment, not balance. Pairs well with L99, but HARDMODE is more aggressive — it refuses to hedge even when hedging would be appropriate.
Example: HARDMODE Which database should I use for my startup: Postgres or MongoDB?
10. /debug
What it does: Traces through your code to find the actual bug, not a list of "common causes".
Why #10: Only works if you paste the actual code + error. Without concrete data, /debug falls back to generic debugging advice. With data, it's the fastest bug-finder I've used.
Example: /debug Here's the failing test and the code under test: [paste both]
11. /trim
What it does: Cuts every word that doesn't change the meaning. Usually trims 30-50% without losing any information.
Why #11: Different from "make it shorter" (which summarizes). /trim preserves every fact but kills filler. Final-pass tool.
Example: /trim [paste your draft]
12. BEASTMODE
What it does: Maximum quality output. Claude takes 3x longer but the result is the kind of work you'd actually ship. Like L99 + /godmode + perfectionism.
Why #12: The "try harder" button. Reserve for things you're actually shipping — landing page copy, important emails, final code that goes to production.
Example: BEASTMODE Write the hero section for a Postgres debugging tool aimed at senior backend engineers.
13. /voice
What it does: Locks a tone for the entire conversation, not just one reply. Specific descriptors work; vague ones don't.
Why #13: Without /voice, Claude's tone drifts back to default after ~5 messages. With it, the tone holds for 30+.
Example: /voice direct, slightly impatient, Stack Overflow top-answer energy
14. /blindspots
What it does: Asks Claude to find what you're missing or assuming, not what you're asking.
Why #14: The most uncomfortable prefix — it finds the premises you didn't realize you were assuming. Uncomfortable but repeatedly useful.
Example: /blindspots Review my plan to migrate from Postgres to MongoDB.
15. FINISH
What it does: Forces Claude to actually complete a task instead of explaining how to do it.
Why #15: Default Claude says "here's how you would do this". FINISH Claude does it. The output is the finished thing — the email, the code, the doc — not instructions for producing it.
Example: FINISH Write the PR description for this diff: [paste diff]
The rest of the 120
These 15 are the highest-use ones. But there are another 105 that serve more specific purposes — formatting codes (/json, /yaml, /table, /checklist), communication codes (/email, /reply, /dm, /escalate), analysis codes (COMPARE, /decision-matrix, /pros-cons, AUDIT), and experimental ones (DOPAMINE, TRACE, /coach, /sage).
The first 11 are free at clskills.in/prompts.
The full set with before/after examples for every code, when-NOT-to-use warnings, code combos, and 10 real-world workflow playbooks is in the Claude Code Cheat Sheet — three tiers from $5 to $10, lifetime updates.
TL;DR
The top 5 in daily use: L99 (committed opinions), PERSONA (expert-level answers), /skeptic (question the question), /ghost (strip AI tells), ULTRATHINK (maximum reasoning depth).
The top 5 for specific tasks: OODA (incidents), /punch (editing), /noyap (quick lookup), HARDMODE (forced decisions), /debug (actual bugs).
The top 5 for polish: /trim (cut filler), BEASTMODE (try harder), /voice (lock tone), /blindspots (find assumptions), FINISH (just do the thing).