25 Secret Commands That 10x Your Claude
25 prompt shortcuts that steer Claude's output. Plus the exact custom-instruction block to paste into a Claude project once so they fire on every message.
By Akhil Boddu · AfricanTechBroWhy commands beat longer prompts
Most people use Claude like a Google search. Type a question, get a wall of hedged, sanitised "As an AI" text, close the tab disappointed.
Power users type a command first. Same model. Different universe of output.
A command is a short token (/ooda, /ghost, /brief) that tells Claude which mode to answer in. You define the modes once inside a Claude project as custom instructions. After that, typing /brief at the top of any message triggers the behaviour. No re-explaining what you want every time.
This guide gives you 25 commands that cover the five problems Claude has out of the box: weak reasoning, AI voice, fluff, blind agreement, and sloppy output format. There's a ready-to-paste custom-instruction block at the bottom. Copy it into your Claude project settings once and the commands work forever.
How to use this guide
- Scroll through the 25 commands. Pick the 5 to 10 that fix problems you actually have.
- Jump to the custom-instruction block at the end.
- Open Claude. Go to Projects. Create a project called something like "Default" or copy it into an existing one.
- Paste the block into the project's custom instructions field.
- Every new chat in that project now understands the commands.
Stack commands by putting two on the same line: /ghost /brief — rewrite this paragraph. Claude applies both.
Category 1: Make Claude think harder
The default Claude answers fast. These slow it down so the reasoning is actually good.
/ooda
What it does: Forces a structured Observe-Orient-Decide-Act pass before Claude writes an answer. Claude lays out what it knows, what the constraints are, what the options are, and which one it's picking. Then it answers.
When to use: Any decision or planning task. Code architecture, deal terms, content strategy, naming things.
Example:
/ooda — should I cold email 100 creators or run a $500 Meta ad to find
my first 10 beta users for a faceless-video SaaS priced at R599/month?
/devil
What it does: Claude argues the opposite side of whatever you said. Kills confirmation bias and surfaces blind spots.
When to use: Before you commit to something. Business decisions, script takes, scripts with a strong claim, negotiation positions.
Example:
/devil — I'm going to raise my freelance rate from R500/hr to R1200/hr
next month. Tell me why this is a bad idea.
/matrix
What it does: Every option laid out as a comparison table with explicit trade-offs. No hand-waving, no "it depends."
When to use: Tool choice, pricing tiers, approach selection. Anywhere "just compare the options properly" is the real ask.
Example:
/matrix — compare Next.js vs Astro vs SvelteKit for a personal blog
with 5000 monthly readers and a single maintainer.
/firstprinciples
What it does: Strips conventional wisdom. Claude reasons from the underlying physics, economics, or human behaviour up to an answer.
When to use: When the industry default feels wrong and you want to verify it actually is. Pricing, product scope, career moves.
Example:
/firstprinciples — is the "ship fast" startup advice actually correct
for a solo founder building a paid SaaS, or is that advice made for
VC-backed teams with a different payoff curve?
/premortem
What it does: Claude assumes your plan failed six months from now and explains exactly why. Then it ranks the failure modes by likelihood.
When to use: Before you commit capital, time, or public reputation. Product launches, campaigns, hiring decisions.
Example:
/premortem — I'm launching a R2000 cohort-based Claude Code course in
60 days with no audience of developers. Imagine it flopped. Why?
Category 2: Fix how Claude sounds
Claude defaults to a hedged, sanitised, slightly corporate voice. These commands rewrite that.
/godmode
What it does: Full creative firepower. No hedging, no disclaimers, no guardrail theatre on topics that don't actually need guardrails.
When to use: Creative writing, hot takes, opinion pieces, persuasive copy. Anywhere default Claude softens your voice into mush.
Example:
/godmode — write a 300-word rant for my newsletter about why most
productivity apps are procrastination in a different outfit.
/ghost
What it does: Rewrites with human rhythm. Shorter sentences. Varied sentence length. Cuts the "In today's rapidly evolving landscape" opener. Cuts the "It's important to note that" filler. Reads like a person wrote it.
When to use: Any text you're about to publish with your name on it. Captions, LinkedIn posts, outreach emails.
Example:
/ghost — rewrite this cover letter so it doesn't read like every
other AI-assisted application on the hiring manager's desk:
[paste cover letter]
Watch out: this is about voice, not detector evasion. AI-detection tools are unreliable and I don't teach detector bypass. If your university requires your own writing, write your own writing and use /ghost to edit.
/artifacts
What it does: Visual and interactive output instead of a text wall. Tables, code blocks, diagrams, side-by-side comparisons.
When to use: Reference material you'll read more than once. Workflows, decision trees, reference tables.
Example:
/artifacts — turn my weekly schedule into a time-blocked table
with colour-coded focus blocks:
[paste schedule]
/eli12
What it does: Explains the topic as if to a bright 12-year-old. Real vocabulary, no condescension. Concrete analogies.
When to use: Complex topic you half-understand. Legal terms, finance, biology, anything you need to explain back to someone else.
Example:
/eli12 — what actually happens when a bank runs out of money, and why
does the government sometimes let it fail and sometimes not?
/expert {field}
What it does: Claude responds as a specialist in the named field. Uses their jargon, their frameworks, their trade-offs.
When to use: When you want depth over accessibility. Code reviews, legal questions, medical research, design critique.
Example:
/expert constitutional lawyer — what are the actual risks of signing
a mutual NDA that has a 5-year term and no carve-out for independent
development?
Category 3: Cut the fluff
Claude likes to pad. These make it stop.
/brief
What it does: Three lines max. Zero preamble. Zero "Great question!" openers.
When to use: Quick lookups, status checks, yes-or-no questions.
Example:
/brief — does Postgres support JSON indexing on nested fields?
/silent
What it does: Just does the task. No explanation, no commentary, no "Here is your result."
When to use: Code generation, rewrites, formatting tasks. When you want output, not a lesson.
Example:
/silent — convert this array of objects into a CSV with headers:
[paste JSON]
/one
What it does: One sentence answer. That's it.
When to use: When you know the question has a simple answer and Claude is about to give you five paragraphs.
Example:
/one — what's the difference between let and const in JavaScript?
/tldr
What it does: Three-bullet summary at the top. Full answer below. You read the top, skip the bottom unless you need it.
When to use: Research questions, long explanations, documentation summaries.
Example:
/tldr — explain what changed between Next.js 14 and Next.js 15 App Router.
/cap {N}
What it does: Caps the response at N words. Claude rewrites internally until it fits.
When to use: Tweet drafts, caption writing, anywhere a length limit matters.
Example:
/cap 220 — write an Instagram caption for a carousel about 3 hidden
features in Claude's Projects feature.
Category 4: Break the yes-machine
Default Claude tries to be helpful. Too helpful. These commands make it argue with you.
/roast
What it does: Brutally honest feedback. Zero sugarcoat. Claude picks apart what's weak.
When to use: Drafts you're about to publish. Pitch decks. Resumes. Scripts.
Example:
/roast — here's my LinkedIn headline. Tell me what's wrong with it
from the perspective of a recruiter skimming 200 profiles a day:
"Building AI-powered solutions at the intersection of tech and impact"
/steal
What it does: Reverse-engineers why something works. Paste a viral post, a great pitch, a good ad, and Claude extracts the structure.
When to use: Any time you see something that worked and want the formula, not a vague copy.
Example:
/steal — this Instagram Reel hit 2.1M views. Extract the exact beat
structure so I can apply the pattern:
"You're not addicted to your phone. You're addicted to avoiding the
silence. Here's what changed when I [...]"
/check
What it does: Claude reviews its own last answer and lists flaws. Forces self-correction.
When to use: After any high-stakes answer. Code that must run, numbers that must be right, claims that will be public.
Example:
/check — review the tax calculation you just gave me. Where could it
be wrong?
/uncertain
What it does: Flags everything Claude isn't sure about. Instead of confident wrong answers, you get "high confidence" and "low confidence" tags.
When to use: Research, medical, legal, anywhere wrong-with-confidence costs you.
Example:
/uncertain — what are the side effects of combining magnesium glycinate
with SSRIs? Flag anything you're not confident about.
/sources
What it does: Every factual claim must include a source or a [verify] tag if Claude doesn't have one. Kills confidently-made-up facts.
When to use: Anything that'll end up in an article, a report, or a public argument.
Example:
/sources — what percentage of South African university students use
ChatGPT for coursework, and how has that changed since 2023?
Category 5: Output format control
Claude's default format is prose. Sometimes you want something else.
/stepbystep
What it does: Numbered instructions only. No prose between steps.
When to use: Tutorials, setup guides, recipes, anything the reader will follow sequentially.
Example:
/stepbystep — how do I set up a Next.js app on Vercel with a custom
domain from GoDaddy?
/json
What it does: Strict JSON output. No markdown, no commentary, no code fences. Just valid JSON.
When to use: Programmatic pipelines. Anywhere you're feeding Claude's output into another script or tool.
Example:
/json — extract the following fields from this email as JSON:
{sender, subject, intent, action_needed, deadline}. Email below:
[paste email]
/outline
What it does: Bullet hierarchy only. No paragraphs.
When to use: Planning a post, structuring a talk, mapping a document before you write it.
Example:
/outline — a 10-minute talk for a student audience on "how to choose
your first programming language in 2026."
/version
What it does: Draft v1. Critique v1. Draft v2. You see the thinking and get the refined output.
When to use: Anywhere the first draft is usually not good enough. Hooks, headlines, subject lines.
Example:
/version — write a subject line for a cold email pitching a
journalist on my AI-proof essay tool. Audience: tech reporters at
TechCrunch, The Verge, Wired.
/tutor
What it does: Claude teaches instead of solving. Asks you to try first, then corrects, then explains.
When to use: When you're trying to LEARN, not just get an answer. Algorithms, maths, languages, debugging.
Example:
/tutor — I'm trying to understand recursion. Give me a simple problem,
let me attempt it, then review my attempt.
The 25 commands at a glance
| # | Command | One-liner |
|---|---|---|
| 1 | /ooda | Structured observe-orient-decide-act before answering |
| 2 | /devil | Argues the opposite side to kill blind spots |
| 3 | /matrix | Comparison table with explicit trade-offs |
| 4 | /firstprinciples | Reason from fundamentals, ignore conventional wisdom |
| 5 | /premortem | Imagine your plan failed, explain why |
| 6 | /godmode | Creative firepower, no hedging, no disclaimers |
| 7 | /ghost | Human rhythm, kills AI voice tells |
| 8 | /artifacts | Visual output, tables, diagrams |
| 9 | /eli12 | Explains like to a bright 12-year-old |
| 10 | /expert {field} | Responds as a domain specialist |
| 11 | /brief | Three lines max, zero preamble |
| 12 | /silent | Just does the task, no commentary |
| 13 | /one | One sentence answer |
| 14 | /tldr | Three-bullet summary at top |
| 15 | /cap {N} | Caps response at N words |
| 16 | /roast | Brutally honest feedback |
| 17 | /steal | Reverse-engineers why something works |
| 18 | /check | Reviews own last answer for flaws |
| 19 | /uncertain | Flags low-confidence claims |
| 20 | /sources | Every fact needs a source or [verify] |
| 21 | /stepbystep | Numbered instructions only |
| 22 | /json | Strict JSON output |
| 23 | /outline | Bullet hierarchy only |
| 24 | /version | Draft, critique, redraft |
| 25 | /tutor | Teaches instead of solving |
Custom-instruction block
Copy everything in the block below. Paste into Claude Projects → your project → Custom instructions.
After this, typing any command at the top of a message triggers the behaviour.
You understand a set of slash commands. When the user's message starts
with one or more commands (tokens beginning with "/"), apply the
corresponding behaviour to your response. Commands can stack on the
same line — apply every one. If no command appears, answer normally.
Commands:
/ooda — Before answering, run a structured observe-orient-decide-act
pass. Lay out what you know, constraints, options, and your pick.
Then give the answer.
/devil — Argue the opposite of what the user said or intends. Your
job is to surface blind spots, not agree.
/matrix — Respond with a comparison table. Columns are the options.
Rows are the trade-offs. No preamble, no conclusion.
/firstprinciples — Strip conventional wisdom. Reason from physics,
economics, or human behaviour up. State the assumptions you're
rejecting.
/premortem — Assume the user's plan failed in 6 months. Explain
exactly why in 3-5 ranked failure modes.
/godmode — Full creative firepower. Drop hedges, drop disclaimers on
topics that don't need them, commit to a voice and opinion.
/ghost — Rewrite with human rhythm. Short sentences. Varied length.
No "In today's evolving landscape", no "It's important to note that",
no "As an AI". Reads like a real person.
/artifacts — Output as visual or structured content: tables, code
blocks, diagrams, side-by-side comparisons. No text walls.
/eli12 — Explain for a bright 12-year-old. Real vocabulary. No
condescension. Concrete analogies.
/expert {field} — Respond as a specialist in {field}. Use their
jargon, frameworks, and trade-offs. State assumptions explicitly.
/brief — Three lines max. No preamble. No "Great question!" openers.
/silent — Just do the task. No explanation, no commentary, no intro
or outro.
/one — One sentence answer. That's it.
/tldr — Three-bullet summary at the top. Full answer below under
heading "Details".
/cap {N} — Cap the response at N words. Rewrite internally until it
fits.
/roast — Brutally honest feedback. Zero sugarcoat. Name what's weak.
/steal — Reverse-engineer why the user's pasted example works.
Extract the structure so the pattern can be applied elsewhere.
/check — Review your own previous answer. List flaws, errors, or
assumptions that could be wrong.
/uncertain — Answer, but tag each claim with [high confidence] or
[low confidence]. Don't hide uncertainty.
/sources — Every factual claim needs a source. If you don't have
one, tag the claim [verify] instead of inventing.
/stepbystep — Numbered instructions only. No prose between steps.
/json — Strict valid JSON output. No markdown, no code fences, no
commentary. Just JSON.
/outline — Bullet hierarchy only. No paragraphs. Max 3 levels deep.
/version — Draft v1. Critique v1 in 3 bullets. Draft v2 addressing
the critique.
/tutor — Teach, don't solve. Give the user a simple version of the
problem. Wait for their attempt. Review. Explain.
When multiple commands are stacked (e.g. "/ghost /brief"), apply all
of them. If two commands conflict, the more restrictive one wins
(e.g. /brief overrides a longer default).
What's next
- Stack commands for compound effects.
/ghost /cap 220gives you human-voice captions under 220 words. - Save the ones you use weekly into a project called "Default" so they fire on every new chat without you pasting the block again.
- Edit the block. Add your own commands. Claude will follow any you define the same way. Good candidates: /brandvoice (your exact voice rules), /icp (your target avatar), /sources only academic.
- Train the habit. For one week, type a command at the top of every Claude chat. After 7 days you'll know which 5 you actually need.
If you've found a command I haven't listed that genuinely changed how you use Claude, DM me on Instagram — I'll add it and credit you.