Someone posts "5 secret Claude codes." A million people watch it. Then they open Claude Code, type the first one, and get nothing back. The commands do not exist. The names were lifted from an old video game with "Claude" stapled to the front. That is the entire trick.
Here is the direct answer. Claude Code ships no built-in /ghost or /godmode command, so typing one into a fresh install does nothing at all. What Claude Code does ship is the ability for you to write your own. A skill is a folder with a SKILL.md file in it, and the folder name becomes the slash command. That is the whole mechanism. So I built the five everyone is pretending to have, as real files you can paste in. They are all in this post.
Nothing here is secret. It is five text files. The reason nobody handed them to you is that nobody had written them.
The Setup, in About a Minute
A Claude Code skill is a directory containing a single SKILL.md file. The directory name is the command: a folder called ghost gives you /ghost. The file needs a small block of YAML frontmatter, a name and a description, and then your instructions in plain Markdown. Anthropic's documentation puts it plainly: "Create a SKILL.md file with instructions, and Claude adds it to its toolkit. Claude uses skills when relevant, or you can invoke one directly with /skill-name."
There are two places a skill can live, and the only difference is reach.
| Level | Where the file goes | Available in |
|---|---|---|
| Personal | ~/.claude/skills/<name>/SKILL.md | All of your projects |
| Project | .claude/skills/<name>/SKILL.md | That one repository only |
Put these five in the personal folder. You will want them everywhere.
mkdir -p ~/.claude/skills/{ghost,uda,buddha,l99,godmode}
Then drop the matching SKILL.md from this post into each folder. Restart Claude Code once. That restart matters only the first time: Claude Code watches skill directories for changes, so once ~/.claude/skills/ exists, editing a skill takes effect in the session you are already in. Creating that top-level directory for the first time is what needs the restart, because the watcher has to pick it up.
The naming is not decoration. The folder name is the command, so keep them exactly: ghost, uda, buddha, l99, godmode. And because each file carries a description telling Claude when to use itself, you do not have to type the slash command at all. Say "make this sound human" and Claude reaches for /ghost on its own.
The Five, at a Glance
| Command | What it does | Reach for it when |
|---|---|---|
/ghost | Rewrites text so it reads like a person wrote it | You want to de-slop your writing |
/uda | Finds the real root cause, not the surface question | Something is failing or stalling |
/buddha | Numbered steps only. No intro, no summary | You want to be told exactly what to do |
/l99 | Talks to you as a peer expert, no dumbing down | You already know the domain |
/godmode | Answers, then argues with itself before shipping | The stakes justify the tokens |
One: Write Like a Human. /ghost.
What it does: rewrites text so a reader cannot tell a model drafted it. It works by deleting the structural tells of LLM prose. The em-dash pivot. The "it is not just A, it is B" contrast. The three-item rhythmic list, which is the single loudest tell of the lot. Throat-clearing openers, summary tails, hedge stacks, and the vocabulary everyone now recognises on sight: delve, leverage, robust, seamless, tapestry.
The important part is what it does after the deleting. Stripping tells gives you clean text. Adding a specific odd detail, an opinion the writer would defend, and a four-word sentence next to a thirty-word one gives you writing. Only the second one sounds like a person.
---
name: ghost
description: Rewrite text so it reads like a person wrote it. Strips the structural tells of LLM prose: em-dash pivots, "it's not X, it's Y", tricolons, hedging, throat-clearing. Use when the user asks to humanize, de-slop, or make writing sound like them.
---
# /ghost: write like a human, not a model
Rewrite the supplied text (or the text you are about to produce) so that a reader
cannot tell a model drafted it.
## Delete these tells, every time
1. **Em-dash pivots.** No `—` used to swing into a reframe. Use a period.
2. **The "not X, but Y" contrast.** Ban `It's not just A, it's B` and every variant.
3. **Tricolons.** Three-item rhythmic lists ("clear, concise, and compelling") are the
single loudest tell. Use one item, or two, or four.
4. **Throat-clearing openers.** No "Great question", "Let's dive in", "In today's
fast-paced world", "At its core".
5. **Summary tails.** No closing paragraph that restates what you just said.
6. **Hedge stacks.** No "can help to potentially improve". Commit to the claim.
7. **Vocabulary flags.** delve, leverage, robust, seamless, tapestry, testament,
landscape, realm, underscore, crucial, pivotal, harness, unlock, elevate.
8. **Uniform sentence length.** Vary hard. A 4-word sentence next to a 30-word one.
## Then add what humans actually do
- One concrete, specific, slightly odd detail. Real writing has texture.
- An opinion the writer would defend, not a balanced survey.
- Contractions. Sentence fragments. The occasional aside.
- Start a sentence with "And" or "But" at least once.
## Voice matching
If the user has given you samples of their own writing (in this conversation, in
CLAUDE.md, or in a linked file), match those first. Their rhythm and vocabulary
override every rule above. If you have no samples, ask for two paragraphs of theirs
before rewriting. Do not invent a voice.
## Output
Return only the rewritten text. No preamble, no explanation of your changes,
unless the user asks for a diff.
## Honest limits
This makes writing *better and more human-sounding*. It is not an "AI detector
bypass". Detectors are unreliable in both directions and any tool claiming a
guaranteed bypass is selling you something. Write things worth reading instead.
Two: Answer the Question Behind the Question. /uda.
What it does: Unpack, Diagnose, Act. Most of the time the question you asked is not the question you needed answered, and a model that answers you literally is being obedient rather than useful. This skill forces three passes. It unpacks what you have actually stated versus what you are quietly assuming. It produces at least three competing hypotheses in a table, weighs them, then commits to one, because a diagnosis that hedges across all three is not a diagnosis. Then it gives you ranked actions ordered by impact divided by effort.
The line that earns its place is in the ACT section. Every action has to come with what result would prove you wrong. That single requirement is what separates advice from a plan.
--- name: uda description: Unpack, Diagnose, Act. Finds the real root cause behind a problem instead of answering the surface question, then gives ranked, testable actions. Use when the user asks why something is failing, stalling, or not working. --- # /uda: Unpack · Diagnose · Act The user asked a question. Your job is to work out whether it is the *right* question, and then answer the one that actually matters. ## 1. UNPACK Restate the situation in one sentence, stripped of the user's framing. Then list, explicitly: - **Given:** the facts stated. - **Assumed:** what the user is treating as true without evidence. Name each one. - **Missing:** the data you would need to be sure. If a missing fact would change your answer, say so and ask for it instead of guessing. ## 2. DIAGNOSE Produce **at least three competing hypotheses** for the root cause. For each: | Hypothesis | Evidence for | Evidence against | Confidence | |---|---|---|---| Then commit. Name the single most likely cause and say why the others lose. A diagnosis that hedges across all three is not a diagnosis. Explicitly separate: - the **surface symptom** (what the user noticed) - the **proximate cause** (what directly produced it) - the **root cause** (the thing that, if fixed, stops it recurring) ## 3. ACT Ranked actions. For each: what to do, what you expect to happen, and **what result would prove you wrong**. Every action must be falsifiable. Order by `impact ÷ effort`, highest first. Cap the list at five. End with the single thing to do in the next hour. ## Rules - No motivational filler. No "you've got this". - If the honest diagnosis is "you don't have enough data yet", say that and specify the cheapest experiment that would get it.
Three: Nothing But the Path. /buddha.
What it does: deletes everything that is not an action. No preamble, no recap, no encouragement, no options. The first character of the reply is the digit 1. Maximum seven steps, every one starting with an imperative verb, every one completable in a single sitting.
The constraint doing the real work is the seven-step cap. If a task genuinely needs more than seven steps, the skill is instructed to treat that as evidence the first step is wrong and to go find a bigger lever. That is not a formatting rule. That is a thinking rule wearing a formatting rule's clothes.
--- name: buddha description: Strip everything that is not an action. Returns numbered, executable steps with zero preamble, zero summary, zero caveats. Use when the user wants to be told exactly what to do, right now. --- # /buddha: nothing but the path Answer with **actions only**. Everything that is not something the user can do is noise, and you delete it. ## Forbidden - Preamble of any kind. Do not restate the question. Do not say what you are about to do. The first character of your reply is the digit `1`. - Closing summaries, recaps, "in short", "to sum up". - Encouragement, empathy, "great question", "you're on the right track". - Caveats and disclaimers, unless the action is genuinely dangerous or irreversible, then one line, inline, in bold. - Options and alternatives. Choose. If you are torn, pick the one you would do and move on. - Explaining *why*, unless the why changes *how* the user executes the step. ## Required shape ``` 1. <verb-first action>: <the one detail needed to execute it> 2. ... ``` - Maximum **7** steps. If it needs more, the first step is wrong; find a bigger lever and start there. - Every step starts with a verb in the imperative. - Every step is completable in one sitting. "Improve your marketing" is not a step. "Rewrite your bio to name one customer and one outcome" is. - No step may depend on information the user has not given you. If it does, make step 1 "find out X" and say precisely where to look. ## The one exception If the user's premise is wrong, if doing any of this would waste their time, say so in a single sentence before step 1. One sentence. Then the steps.
Four: The Training Wheels Come Off. /l99.
What it does: treats you as a peer expert. No opening definition, no analogy, no "simply" or "basically", no beginner path running alongside the real one. It starts where the obvious approach stops working, because you already know the obvious approach.
Two rules make it worth having. First, give numbers: "slower" is useless, "roughly 40 times slower above 10,000 rows" is an answer. Second, state your confidence and say what would move it. An expert can handle "I am 60% on this, and here is what would settle it." At this level a fabricated detail is far worse than an admitted gap, and the file says so explicitly.
--- name: l99 description: Talk to the user as a peer expert. No simplifying, no analogies, no defining terms, no beginner scaffolding. Use when the user already knows the domain and wants the real answer at full resolution. --- # /l99: the training wheels come off The user is an expert. Treat them as one. Every sentence you spend explaining something they already know is a sentence you stole from them. ## Assume the user knows - The domain's standard vocabulary. Use the real terms. Do not gloss them. - The obvious approach, and why it is obvious. Skip it. Start where it stops working. - The textbook answer. They want the part the textbook leaves out. ## Therefore, do not - Open with a definition. Ever. - Use an analogy. If a concept only survives as a metaphor, you do not understand it well enough to be talking about it at this level. - Write "simply", "just", "basically", "essentially". - Offer a beginner path alongside the real one. - Soften a technical disagreement into a preference. ## Instead, do - **Lead with the non-obvious.** The first line should be the thing they could not have got from the first page of search results. - **Name the tradeoff.** Every real answer at this level is a tradeoff. State both sides and where the crossover point sits. - **Cite the failure mode.** What breaks, at what scale, and what the symptom looks like when it does. - **Give numbers.** Orders of magnitude, complexity classes, actual thresholds. "Slower" is useless. "~40× slower above 10k rows" is an answer. - **Disagree.** If the user's stated approach is wrong, open with that. ## Calibration State your confidence explicitly where it is below certain, and say what would move it. An expert can handle "I'm 60% on this, and here's what would settle it." Never bluff. At this level a fabricated detail is worse than an admitted gap.
Five: The One That Argues With Itself. /godmode.
What it does: the expensive one. Four passes. It frames the real question, answers it at peer-expert resolution, then red-teams its own answer, then ships the version that survived. Pass 3 is the whole point: find the load-bearing claim most likely to be false, name the strongest counter-argument a smart opponent would open with, and say what context would flip the recommendation.
The instruction I am proudest of is the one that makes the pass honest. "A red-team pass that concludes 'the answer holds up well' has failed." And if Pass 3 destroys Pass 2, the skill ships the reversal. Changing its mind mid-answer is the feature, not the bug. It closes with the call, a confidence number, and the tell: the first observable sign that the call was wrong, so you know when to abandon it.
--- name: godmode description: Maximum-depth answer. Runs the full chain: diagnose, answer at expert level, then red-team its own answer and ship the surviving version. Use when the stakes are high enough to justify the tokens. --- # /godmode: the strongest version you can produce This is the expensive one. The user has decided the answer matters more than the latency. Spend accordingly. ## Pass 1: Frame State the real question (see `/uda` if the framing is suspect). Name the decision this answer feeds into, and the cost of getting it wrong. If the stakes are low, say so and downgrade. Do not burn a thousand tokens on a coin flip. ## Pass 2: Answer Answer at peer-expert resolution (see `/l99`). Full depth. Numbers, tradeoffs, failure modes, thresholds. No simplifying. ## Pass 3: Red-team yourself This is the pass that makes the skill worth running. Attack your own Pass 2: - Where is it **wrong**? Find the load-bearing claim most likely to be false. - Where is it **thin**? What did you assert with confidence you have not earned? - What is the **strongest counter-argument** a smart opponent would open with? - What **context would flip the recommendation**? Name it explicitly. Be genuinely adversarial. A red-team pass that concludes "the answer holds up well" has failed. Find the crack. ## Pass 4: Ship Rewrite the answer with the red-team's findings folded in. Then close with: - **The call:** one line, no hedging. - **Confidence:** a number, and what would move it. - **The tell:** the first observable sign that this call was wrong, so the user knows when to abandon it. ## Rules - Never pad to seem thorough. Depth is resolution, not word count. - If Pass 3 destroys Pass 2, ship the reversal. Changing your mind mid-answer is the feature, not a bug.
The Honest Part
Three things, and the first one is the one people will not want to hear.
/ghost is not an AI-detector bypass. It makes writing sound like a person because it makes writing better, and better writing happens to read as human. That is the mechanism. It is not a cloak. Detectors are unreliable in both directions, they flag real human writing and they miss real model output, and anyone selling you a guaranteed bypass is selling you something. I put that limitation inside the skill file itself so the model repeats it back to you rather than letting you believe otherwise.
/godmode genuinely costs more. Four passes means four passes of tokens. It is worth it for a pricing decision or an architecture call. It is not worth it for renaming a file, and the skill is written to tell you so and downgrade itself when the stakes are low.
These are prompts, not superpowers. A skill is a text file that steers a model you already had. It does not unlock a hidden mode, because there is no hidden mode. What it does is stop you retyping the same four paragraphs of instruction every session, and make the good version of your instruction the default instead of the version you could be bothered to type at 11pm.
The five secret codes were never secret. They were never codes. They are five text files, and now they are yours.
What You Actually Just Got
- 01/ghost for writing that does not announce its author. Deletes the tells, then adds the texture.
- 02/uda for the diagnosis you needed rather than the question you asked. Three hypotheses, one commitment.
- 03/buddha for seven steps and not one word more. The reply opens with the digit 1.
- 04/l99 for the answer at full resolution, with numbers and a confidence level attached.
- 05/godmode for when being right matters more than being fast. It attacks its own answer before you see it.
Once these are in, the natural next move is to stop writing skills one at a time and start installing other people's. That is what the Claude Code plugin ecosystem is, and it is also how you give Claude Code a memory that survives the session. And if you want it to stop burning through your usage limit while it does all this, the fourteen commands that keep Claude Code at its desk are the other half of this setup.
Everybody else is posting fake commands. Build the stuff they are faking.
Questions People Ask
Are /ghost and /godmode real Claude Code commands?
No. Claude Code ships no built-in /ghost or /godmode, so typing one into a fresh install does nothing. They become real only when you write them as custom skills, which is exactly what the files in this post are.
How do I create a custom slash command in Claude Code?
Create a folder inside ~/.claude/skills/ and put a SKILL.md file in it. The folder name becomes the command, so ~/.claude/skills/ghost/SKILL.md gives you /ghost. The file needs frontmatter with a name and a description, then your instructions in Markdown.
Where do Claude Code skills live?
Two levels. Personal skills sit at ~/.claude/skills/<name>/SKILL.md and work across all your projects. Project skills sit at .claude/skills/<name>/SKILL.md and work in that repository only. When names clash, the personal skill wins.
What is the difference between a skill and a custom command?
They are the same thing now. Anthropic's docs state that custom commands have been merged into skills: .claude/commands/deploy.md and .claude/skills/deploy/SKILL.md both create /deploy. Skills add a folder for supporting files and let Claude invoke them automatically when they are relevant.
Can a skill bypass an AI detector?
No, and no tool reliably can. /ghost strips the structural tells of model prose, which makes text read as human because it is better writing. Detectors guess, and they guess wrong in both directions. Any product promising a guaranteed bypass is selling you something.
Do I have to restart Claude Code after adding a skill?
Only the first time. Claude Code watches skill directories, so editing a skill inside an existing ~/.claude/skills/ takes effect immediately. Creating that top-level directory when it did not exist at session start is what requires the restart.
Want an AI setup that works the way you actually work?
I help founders and brands build generative AI systems that ship: agentic workflows, coding-agent setups, and content engines that are production-grade down to the plumbing. If you want a second pair of eyes on yours, book a session.
Book a Growth Chat
Skill file format, directory locations, the merge of custom commands into skills, and live change detection were checked against Anthropic's Claude Code documentation on skills (code.claude.com/docs/en/skills) and slash commands (code.claude.com/docs/en/commands), current as of July 2026. Claude Code skills follow the Agent Skills open standard, so the same SKILL.md format carries to other tools that support it. The five skill files in this post are my own work and are free to copy, edit, and rename. Nothing in this post is sponsored.