AI Guides › Workbench
By Nigel Guy · 6 min read
Lists of "secret Claude commands" tend to mix two different things. Some are real switches that the software recognises and acts on. Others are ordinary sentences that happen to work because they are clearer than what you would have typed. Both are useful, but only one kind can be looked up in the documentation, and confusing them is how people end up typing magic words into a box that ignores them.
The rule: know which of your seven are switches and which are instructions, and only expect a switch to behave the same way every time.
This kit has three switches (all from Claude Code, Anthropic's coding tool, checked against its current docs) and four instruction patterns that work in any Claude chat. A note on provenance: this topic came from a social post claiming a much longer list. We could not verify that list, so we have kept only what we could check.
| # | Name | Kind | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|---|---|
| 1 | ultrathink |
Real keyword (Claude Code) | Asks for deeper reasoning on that one turn | Claude Code needs a paid plan; Pro is $20 a month (about £15 at time of writing, check the £ price at checkout) | Hard bugs, design choices | Does not change your effort setting; nothing else like it is recognised |
| 2 | /plan |
Real command (Claude Code) | Enters plan mode: Claude reads and proposes, no edits until you approve | Same plan | Anything touching many files | You must read the plan, not just approve it |
| 3 | /compact [focus] |
Real command (Claude Code) | Summarises the conversation to free context; your focus text steers the summary | Same plan | Long sessions drifting off track | Detail you did not name can be lost |
| 4 | Interview me first | Instruction | Makes Claude ask for missing inputs before starting | Free in any Claude plan | Vague briefs | Slows quick tasks |
| 5 | Options with trade-offs | Instruction | Forces comparison instead of one confident answer | Free | Decisions | Only as good as the criteria you give |
| 6 | Self-check before answering | Instruction | Claude tests its draft against your criteria | Free | Anything with a checklist | A check is not proof |
| 7 | Argue against it | Instruction | Claude attacks its own recommendation | Free | Before committing money or time | It will find faults even in good plans |
ultrathinkAnthropic's model configuration page says that including ultrathink anywhere in your prompt requests deeper reasoning on that turn, without changing your session's effort setting. It also says other phrases such as "think hard" or "think more" are passed through as ordinary text and are not recognised as keywords.
How to use it: put the word in a prompt where the problem is genuinely hard.
ultrathink. Here is a bug I cannot explain: [DESCRIBE_SYMPTOM_AND_STEPS_TO_REPRODUCE].
Relevant files: [FILE_PATHS]. List the three most likely causes in order of
probability, say what evidence would confirm each, and ask me for anything you
need before guessing.
Fill in the symptom, the steps and the files. For a standing change, use /effort instead: with no argument it opens a slider, and /effort plus a level name sets it directly. The levels available depend on the model; the current docs list low, medium, high, xhigh and max for the newest models.
/planTyping /plan enters plan mode directly; /plan fix the auth bug enters it and starts on that task. In plan mode Claude reads files and proposes a plan but makes no edits until you approve. You can also press Shift+Tab until the status bar shows plan mode is on.
Use it before any change you would not want to undo by hand. Read the proposal for files it plans to touch that you did not expect.
/compact with a focus/compact summarises the conversation so far to free up room. You can add instructions, so say what must survive:
/compact Keep: the decisions we made, the file names we changed, and the failing
test output. Drop: the exploratory dead ends.
If you want a clean slate instead, /clear starts a new conversation (the old one stays available through /resume). Compacting continues the same conversation; clearing does not.
These are not switches. Type them into any Claude chat, and they work because they are specific.
Act as a [ROLE, e.g. careful project editor]. My goal: [GOAL]. A good result
looks like: [WHAT_DONE_LOOKS_LIKE]. Before you write anything, ask me up to
[NUMBER] questions about missing inputs, audience, limits and format. Do not
guess any answer. Once I reply, produce [OUTPUT] in [FORMAT].
I must decide: [DECISION]. Context: [CONTEXT]. My constraints: [BUDGET, DEADLINE,
SKILLS]. Give me three genuinely different options as a table with columns:
what it is, cost, main risk, what would make it the wrong choice. Then
recommend one and say what you assumed. If you lack a fact, ask me.
Task: [TASK]. Success criteria: [CRITERIA_LIST]. Draft your answer, then
check it against each criterion one by one. Fix anything that fails. Show me
only the final answer plus a one-line pass/fail per criterion. If a criterion
cannot be checked from what I gave you, say so.
Here is my plan: [PLAN]. Act as a sceptical reviewer who wants me to succeed.
List the five strongest objections, ranked by how much damage they would do.
For each, say what evidence would settle it and a cheaper way to test it. Do
not soften the objections, and do not invent facts about my situation.
The trap is treating phrases as incantations. "Think harder", "be smarter" and "expert mode" are not recognised switches in Claude Code, and in an ordinary chat they are just vague instructions. Only ultrathink is named in the documentation as a keyword. Anything a blog calls a "secret code" should be checked against the commands reference, and if it is not there, treat it as a plain sentence and judge it by whether it makes your request clearer.
A second trap is volatility. Claude Code's commands and effort levels change between releases; run / in your own session to see what you actually have.
Because the useful ones share two jobs: control how much thinking and context Claude uses (switches 1 to 3), and force it to be specific about inputs, options, checks and objections (4 to 7). A list of seventy would be mostly rewordings of those. Memorise the three switches, keep the four prompts in a notes file, and stop collecting.
ultrathink./plan./compact with a focus, or /clear.ultrathink on every message. It is for the hard turn, and a one-line question does not need it.