AI Guides › Workbench
By Nigel Guy · 8 min read
You see a post promising "five secret Claude slash commands", type /redteam into a chat, and get something useful back. So you assume you've found a hidden feature. You haven't. You've written a short instruction with a slash in front of it, and Claude read it like any other text. That works until you rely on it somewhere it doesn't.
The rule: a slash in front of a word doesn't give it any power. What makes the answer change is the instruction you write after it. Write that instruction out in full, and only turn it into a real command when you've used it often enough to justify it.
Mostly no. Here's how it stands as of 4 October 2026:
/human, /mirror, /redteam and /forecast aren't documented commands. Claude reads them as plain words and guesses what you mean./output-style concise, or run /config and choose Output style. It needs v2.1.269 or later for the command, and v2.1.237 or later for the style itself..claude/skills/redteam/SKILL.md (or the older .claude/commands/redteam.md) creates a working /redteam command. In the Claude apps, you can upload a custom skill and turn it on under Customize > Skills.| Mode | What it forces | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| Red team | Claude attacks your plan instead of improving it | £0. It's a prompt | Anything you're about to ship, sign or send | It only attacks the details you give it |
| Forecast | Ranked outcomes with what would cause each one | £0 | Go/no-go decisions | It reads as confident when it isn't. Ask for the uncertainty |
| Mirror | Plays your own words back, including the gaps | £0 | Fuzzy briefs and goals you can't explain clearly | Feels slow when you want an answer |
| Human | Plain speech, no list-heavy formatting | £0 | Messages to real people | It can't add warmth that isn't in your draft |
| Concise | Answer first, no padding | £0 as a prompt. A built-in style in Claude Code | Status checks, quick facts | It can drop caveats you needed |
Skills are available on Free, Pro, Max, Team and Enterprise plans and need code execution switched on. Anthropic lists Pro at US$20 a month billed monthly; check the sterling price at checkout.
A review gets you polite suggestions; a red team gets you the failure point. Use it before anything you can't easily reverse.
You are a sceptical reviewer whose only job is to find what breaks. Do not encourage me or improve the plan.
My plan: [THE SPECIFIC PLAN — who, what, when, how much, not a category like "launch a product"]
What I'm assuming: [YOUR KEY ASSUMPTIONS, OR "unsure"]
What failure would cost me: [MONEY, TIME, REPUTATION, OR OTHER]
Steps:
1. If the plan is too vague to attack, ask me up to three questions and stop.
2. Name the single weakest assumption and explain why it is the weakest.
3. Describe the first thing that goes wrong if that assumption is false, as a short sequence of events.
4. List two further risks, ranked.
5. For each risk, give one cheap test I could run this week to check it.
Output: a short heading for each step, plain sentences, under 350 words.
Rules: label every point as FACT (from what I told you) or GUESS (your inference). Do not invent figures about my market or customers.
Before answering, check: did I attack the plan as written, or a generic version of it?
Fill in: your real plan, your assumptions and what failure would cost you.
Act as a forecaster who is careful about uncertainty.
Decision: [WHAT YOU ARE ABOUT TO DO]
Timeframe: [E.G. NEXT 90 DAYS]
Known facts: [WHAT YOU ACTUALLY KNOW]
Give the three most likely outcomes if I go ahead unchanged. For each one: a one-line description, the main thing that would cause it, an early warning sign I could watch for, and your confidence as low, medium or high, with a reason. Do not give percentages unless I have supplied data that supports them. If a missing fact would change the ranking, name it and ask for it instead of assuming.
Check before answering: are these outcomes different from each other, or three versions of the same one?
Fill in: the decision, the timeframe and the facts you have.
Before you help, restate my request back to me.
My request: [WHAT YOU WANT]
1. Summarise in two sentences what you think I'm asking for and why.
2. List anything I left unstated that you would otherwise have to assume.
3. Point out any contradiction in what I wrote.
4. Ask me to confirm or correct before you do any work.
Do not start the task in this reply.
Fill in: the request you're not sure you've explained well.
Rewrite the text below so it reads like a person wrote it to another person.
Text: [YOUR DRAFT]
Reader: [WHO RECEIVES IT AND YOUR RELATIONSHIP TO THEM]
Channel: [EMAIL, TEXT, SLACK, LETTER]
Keep every fact and commitment the same. Use ordinary sentences, not bullet points or headings, unless the original had them. Remove stock phrases and filler openings. Do not add warmth, apologies or promises I didn't make. If the tone I want is unclear, ask me.
Check: did any fact, date or figure change? If so, fix it.
Fill in: your draft, who reads it and where it's going.
In a chat, add this to the end of any question:
Answer in the fewest sentences that are still correct. Lead with the answer. Keep any warning I need in order to act safely, even if it adds a sentence. If the honest answer is "it depends", say what it depends on in one line.
In Claude Code, use the built-in style (/output-style concise) instead. It's maintained by Anthropic, and it keeps error reports and security warnings at full length.
Order matters: mirror → red team → forecast → concise. Mirror confirms the right plan, forecast then accounts for the weak point, and concise summarises the full analysis rather than replacing it. Keep Human for text going to another person. Once the brief is confirmed, use this:
My plan, stated precisely: [PLAN]
Part 1 — Attack it: name the weakest assumption and what fails first if it's wrong.
Part 2 — Project it: three distinct likely outcomes if I proceed unchanged, each with an early warning sign.
Part 3 — Compress it: one sentence I most need to hear before deciding.
Mark anything you're inferring about my situation as GUESS. Ask for missing facts rather than filling them in.
Claude Code now lets you chain skills at the start of a message, for example /skill-a /skill-b do XYZ. You can chain up to six, and it needs v2.1.199 or later. If you've built red team and forecast as skills, you can call them together.
You run a small candle business and plan a 15% price rise. Mirror shows you never mentioned wholesale customers. Red team names the weakest assumption (repeat buyers won't switch) and the first break (wholesale reorders). Forecast's early warning sign is basket size in the first fortnight. Concise: test the price on retail before touching wholesale. This illustrates the stack, not real candle shops.