What it is. Drop the single word ultrathink anywhere in a prompt and Claude Code spends a deeper reasoning pass on that turn. Per Anthropic's own docs: "Include ultrathink anywhere in your prompt to request deeper reasoning on that turn without changing your session effort setting. Claude Code recognizes the keyword and adds an in-context instruction."
The honest correction (this is where the viral reels lie). "think", "think hard" and "think more" are NOT magic keywords. The same doc says they "are passed through as ordinary prompt text and are not recognized as keywords." So the old "think / think hard / ultrathink are escalating tiers" claim is wrong. Only ultrathink is a real trigger, and it adds an instruction rather than changing the effort level sent to the API.
When to use it. A gnarly bug, an architecture decision, or a plan you want Claude to reason through harder than a normal reply. For a persistent setting instead of a one-off, use /effort (low, medium, high, xhigh) which is the actual control for how much Claude thinks.
Copy this:
ultrathink, review this function for edge cases before you rewrite it, and list the risks you find first.
Source: Claude Code model configuration docs
What it is. A skill is a SKILL.md file with instructions that Claude can run on demand as /skill-name, or load automatically when relevant. You can write your own or install ones other people share. Claude Code skills follow the open Agent Skills standard.
Two real ways to add one:
~/.claude/skills/terse/SKILL.md with instructions like "answer in the fewest words possible, no preamble, no summaries", then invoke it with /terse. It loads only when used, so it costs almost nothing until you run it./plugin marketplace add anthropics/claude-plugins-community, then /plugin install <plugin-name>@claude-community, then /reload-plugins to activate it.When to use it. The token-saving idea from the reel is real: a skill that forces short, terse output does cut how many tokens Claude spends on giant replies. But the honest version is you make (or install) that behavior yourself, and the actual number saved depends entirely on your prompts. I could not verify any specific "cut 75% of tokens instantly" community skill against a primary source, so treat that figure as marketing, not fact. The mechanism is legit, the exact percentage is not.
Source: Extend Claude with skills · Discover and install plugins
What it is. A built-in Claude Code command. The commands reference describes it as: "Generate a report analyzing your Claude Code sessions, including project areas, interaction patterns, and friction points." It reads your local session history and hands back where you are losing time.
When to use it. Once a week or once a month, to see your recurring patterns, your friction points, and the instructions you keep repeating (which are exactly the things worth turning into a CLAUDE.md rule or a skill).
Copy this:
/insights
/usage for a token and cost breakdown of your current session.Source: Claude Code commands reference · Manage costs (/usage)
What it is. /loop is a bundled skill that reruns a prompt on a repeat while your session stays open, so you stop babysitting a build or a PR. Give it an interval and a prompt and Claude converts it to a cron job and confirms the cadence.
Copy this:
/loop 5m check if the deploy finished and tell me what happened
When to use it. Polling a deployment, watching CI, babysitting a long build, or a one-off reminder ("remind me at 3pm to push the release branch").
/loop tasks are session-scoped: they only fire while Claude Code is running and idle, they stop when you start a fresh conversation, and recurring tasks auto-expire 7 days after creation. For scheduling that survives without an open session, use Routines (Anthropic cloud), Desktop scheduled tasks, or GitHub Actions instead.Source: Run prompts on a schedule
What it is. The genuinely underused one. While Claude is mid-task, you do NOT have to stop it to add a thought. Type your message and press Enter: the running tool finishes, and Claude reads your note and folds it in at its next step. Only pressing Esc actually interrupts and stops the work in flight.
How it works under the hood. Per Anthropic's docs, a queued prompt "fires between your turns, not while Claude is mid-response. If Claude is busy when a task comes due, the prompt waits until the current turn ends." So you get to steer without restarting.
The honest correction. The viral reel calls this a BTW command. There is no BTW command. The real behavior is just: type and press Enter to queue, press Esc to interrupt. That is it.
When to use it. You are watching Claude build and you realize you forgot a requirement. Instead of Esc-ing and losing the in-flight work, you queue "also add input validation" and it gets picked up cleanly.
Source: Run prompts on a schedule (how scheduled tasks run)
ultrathink, before you write any code, map the 3 riskiest assumptions in this plan and how you would test each one.
Look back over how I have used Claude Code in this project and give me one skill and one CLAUDE.md rule that would save me the most time, based on what I keep repeating.
/loop 10m check whether CI passed on the current branch and address any new review comments, then tell me what you changed.
If you are newer: start with just two of these, ultrathink and the type-and-Enter queue trick. They need zero setup and they change how a single session feels immediately. Type /help in Claude Code any time to see the live list of commands your version has.
If you are a power user: the compounding move is #2 and #3 together. Run /insights to find the instructions you repeat most, then bake each one into a SKILL.md or a CLAUDE.md rule so you never type it again. The point of these is not that you memorize five secrets, it is that you build a setup once and reuse it forever.
claude --version to see what you are on, and check the Claude Code docs if something here behaves differently. Everything above was verified against Anthropic's live docs, but the docs are the source of truth, not any reel (including this one).