AI Guides › Workbench
By Nigel Guy · 8 min read
A post promises that typing /truth or /eli5 before a prompt unlocks a hidden Claude mode. You try it, the reply seems a little different, so you keep using it. The difference was Claude guessing what one word meant. Sometimes it guesses well. Sometimes it gives you a vague version of what you wanted, and nothing tells you which happened.
The rule: a label is a shortcut to an instruction you've already written down. Write the instruction once, in a legend Claude can see, then use the short label as often as you like.
Anthropic hasn't documented any built-in label like /truth or /humanize for Claude chat. In ordinary chat, a slash word is just text, and Claude reads it the way it reads anything else. Anthropic's prompting guide compares Claude to "a brilliant but new employee who lacks context on your norms", and a new employee would have to guess what "/stress" means.
Slashes do mean something in two places, and you can use both deliberately:
/ opens a menu of built-in commands and skills. A personal skill saved at ~/.claude/skills/<name>/SKILL.md becomes /<name>./ in the sidebar lists your skills.If a label shares a name with a skill or command you've installed, rename the label.
The labels are typed text, so they cost £0 and work on every plan, including Free. Here are fifteen, grouped by the job they do.
| Label | What it does | Best for | Catch |
|---|---|---|---|
/human |
Plain, varied sentences. No stock phrases, no tidy three-part lists | Emails, posts, bios | Can't make up for having nothing to say |
/truth |
Marks each claim as checked, inferred or guessed, and says "I don't know" when that's the answer | Facts, figures, dates | Tags are Claude's self-assessment, not proof (see below) |
/stress |
Attacks the plan: weakest assumption, cheapest test, what breaks first | Plans, pricing, launches | Can invent objections to fill the slot |
/reframe |
Rewrites your question into a sharper one before answering | Vague or loaded questions | Can drift from what you actually needed |
/options |
Several genuinely different approaches, with trade-offs | Early-stage decisions | More choice isn't always better. Cap it at three |
/next |
One concrete action you can take in the next hour | Stuck projects | Can over-simplify a real dependency |
/askme |
Asks you questions before answering | Briefs, plans, anything personal | Slower. Skip it for lookups |
/simple |
Explains for a newcomer, with one analogy | New topics | Analogies leak. Check the edge cases |
/short |
Answer first, five lines or fewer | Quick checks | Cuts the reasoning you might need |
/premortem |
Assumes it failed in six months, then explains why | Commitments you can't easily undo | Story-shaped, not a probability |
/root |
Keeps asking "why" until it reaches a cause you control | Recurring problems | Can stop at a cause that just sounds plausible |
/reader |
Rewrites for a named reader and what they need to decide | Reports, pitches | Needs you to name the reader |
/redline |
Line edits, each with a reason, ranked by impact | Your own drafts | Can flatten your voice. Reject freely |
/debate |
Strongest case for, strongest case against, then a verdict | Arguments you're attached to | The verdict is still Claude's judgement |
/decide |
Picks one option and states what would change the pick | When you've stopped learning anything new | Don't use it on choices only you can weigh |
Paste this at the start of a chat, or save it (Step 2). Delete labels you won't use; a shorter legend is easier for both of you to follow.
You are my working partner. I sometimes start or end a message with a label from the legend below. When I do, follow that label's instruction on top of my request. Labels can be combined. If two of them conflict, tell me which one you followed and why.
/human - Varied sentences, no stock phrases, no automatic lists of three, no closing summary.
/truth - Tag factual claims [checked], [inferred] or [guess]. Say "I don't know" rather than fill gaps.
/stress - First name the weakest assumption, what breaks first, and the cheapest test.
/reframe - Show me the more useful question under mine, then answer that.
/options - Up to three approaches that differ in kind, each with its main trade-off.
/next - End with one action I can finish within an hour.
/askme - Ask me up to five questions first. Don't answer until I reply.
/simple - Explain for a newcomer with one analogy, and say where it breaks.
/short - Put the answer in the first line. Five lines at most.
/premortem - Assume this failed six months from now. List the likeliest reasons first.
/root - Ask "why" until you reach a cause I can act on. Show each step.
/reader - Rewrite for [READER] and the decision they need to make.
/redline - Line edits with one-line reasons, ranked by impact. Keep my voice.
/debate - Strongest case for, strongest case against, then a verdict with your confidence.
/decide - Pick one option and say what would change your pick.
If a label needs information I haven't given you (for example /reader with no reader named), ask me for it rather than guessing. Before you send any labelled reply, check that every label I used has been applied.
What to fill in: nothing until you use /reader. Then name the reader in your message, for example "/reader the finance director".
| Where | How (at time of writing) | Scope | Watch out for |
|---|---|---|---|
| Instructions for Claude | Your initials (bottom left) > Settings > Instructions for Claude | Every conversation | Also applies where you don't want it |
| Project instructions | Inside a project, in its instructions | Chats in that project only | Free accounts are limited to five projects |
| Custom skill | Upload a ZIP under Customize > Skills | Loads when relevant, or when you invoke it | Needs code execution switched on |
| Claude Code skill | ~/.claude/skills/<label>/SKILL.md |
Every project, as /<label> |
One folder per label |
For a label you use every day, a Claude Code skill makes it a real command. Add disable-model-invocation: true so it only runs when you type it, and use $ARGUMENTS to pass in whatever you type after the name:
---
name: premortem
description: Failure analysis of a plan before committing to it
disable-model-invocation: true
---
You are reviewing a plan I'm about to commit to: $ARGUMENTS
1. If the plan, deadline or budget is missing, ask for it and stop.
2. Assume it failed six months from now. Give the five likeliest causes as a table: cause, early warning sign, cheap prevention.
3. Before replying, delete any cause that would fit almost any plan.
/askmeMost bad replies come from missing context, not from weak wording. /askme makes the gap visible before Claude fills it with assumptions.
/askme I need a one-page brief for [PROJECT] going to [AUDIENCE] by [DATE].
Fill in the project, audience and deadline. Answer the questions it asks, then let it write.
/truth more than it deserves/truth borrows a technique Anthropic recommends against hallucinations: let Claude say "I don't know". It helps, but a [checked] tag is still Claude describing its own confidence. Unless web search or a document you supplied backs the claim, nothing was checked. Use the tags to decide what to verify first, and verify anything you'll publish, sign or pay for.
| If your problem is... | Use |
|---|---|
| The reply sounds like a machine wrote it | /human, /redline |
| You don't trust the answer | /truth, /debate |
| You're about to commit | /stress, /premortem, /decide |
| You're not sure what to ask | /askme, /reframe, /options |
| The answer misses the person reading it | /reader, /simple, /short |
| The same problem keeps coming back | /root, /next |
Start with three labels from three different rows, for example /askme, /stress and /short, and use only those for a week. You'll learn which ones change the answers and which you add out of habit.
/short on anything with a risk attached. The reasoning it cuts is the part you need to check./ in the main Claude web chat opens a skills menu. Anthropic documents that menu for Claude Code and the Microsoft 365 add-ins only.