Claude Secret Prompt Codes
100 prompt-engineering shortcuts
This is the full set of 100 Claude codes I've been collecting and testing for over a year. Organized into 10 categories by what you're trying to do: voice, build, think, learn, strategize, create content, code, research, power-shape the conversation, and operator-mode the response. Most of these codes are not magic. They're prompt-engineering shortcuts that work because Claude has been trained on the patterns the words point to. The ones with established public definitions (OODA, MECE, FACTCHECK) work consistently. The made-up ones (/ghost, /godmode, BEASTMODE) work most of the time, but treat them like vocabulary, not spells. If a code doesn't work the first time, write the instruction in plain language instead.
Most prompt-engineering content treats Claude codes like secret incantations. Paste this magic word, get a magic result. That's not how this works, and pretending otherwise is how you end up confused when your perfect prompt returns generic slop.
Here's the actual truth, up front, because it makes everything else useful:
These codes are vocabulary, not magic.
They work because Claude has been trained on the patterns the words point to. /ghost works because Claude has seen "make this sound human, remove AI tells" a thousand different ways in its training data. OODA works because the Observe-Orient-Decide-Act framework is documented in actual military strategy texts. /godmode works because "give me your most comprehensive response, drop the brevity preference" is a request Claude understands.
That has three consequences:
- The codes with established public definitions work consistently. OODA, MECE, BLUF, FACTCHECK, SWOT, [JSON], [TABLE]. These have clear meanings Claude has been trained on. They land almost every time.
- The made-up codes work most of the time, not always. /ghost, /godmode, BEASTMODE, CEOMODE. These rely on Claude inferring intent from context. Usually they hit. Sometimes they don't.
- The format codes work on every model. [JSON], [TABLE], [BULLETS]. Format instructions are the most reliable shorthand on this list. They work because the output format is unambiguous.
The rest of this prompt pack is below.
Enter your email to read it. You also get one new playbook every week.
One email a week. Unsubscribe any time.
If a code doesn't return what you expected, the issue usually isn't the code. It's missing context. Add a sentence about the audience, the use case, or the outcome you want. The code does the heavy lifting on register and structure. You still have to give it something specific to work with.
This is the full set of 100 codes I've been collecting over the last year, organized by what you're actually trying to do. Some I use every day. Some I tested once and saved for the right moment. Some I'm including because operators in my watchlist get good results from them and they're worth running through your own workflow.
How to use the rest of the list
Each code is a one-line instruction. Paste it before your prompt unless you've got a feel for where it lands better. Some belong at the end (L99, [JSON], [TABLE]). Most work fine at the start.
If a code returns something flat or generic, the issue isn't the code. It's that Claude needs more context. Add a sentence about the audience, the use case, or the outcome you want. The code does the heavy lifting on register and structure. You still have to give it something specific to work with.
The 5 codes from the reel
These are the headliners. The rest of the list builds on these patterns.
Voice & Style (10 codes)
Most people get the most mileage out of these. Voice fixes are the fastest visible improvement to AI output.
Build & Ship (10 codes)
Where the actual artifacts feature lives plus the prompt patterns that get the most out of it.
Decision-Grade Thinking (10 codes)
These are the codes I lean on when the answer matters more than the speed. Frameworks Claude has been trained on actually do work better than generic "think harder" prompts.
Skill Acceleration (10 codes)
For when you're trying to learn something fast, not just answer a question.
Strategy & Analysis (10 codes)
The decision-support set. Most of these I run during scoping calls or before locking on a direction.
Content & Hooks (10 codes)
The set most relevant to anyone building an audience. I use four of these every week inside the Content Studio.
Code & Infrastructure (10 codes)
If you write code or work with engineers, this is the set. I lean on /shipit and ARCHITECT regularly when scoping technical work.
Research & Intel (10 codes)
For when you need to know something fast and need it accurate.
Power Commands (10 codes)
The meta-codes. These shape how Claude responds across the rest of the conversation, not just the next answer.
Operator Modes (10 codes)
The ones operators in my watchlist use most. Some of these I haven't personally tested but the pattern is solid.
