The 10x
Operator List

Run three sessions in three git worktrees: one builds the feature, one fixes the bug, one refactors. You review and merge. The bottleneck stops being typing speed and becomes your judgment, which is the job anyway.

Point it at a repo you've never seen and ask where the auth flow lives, why a request is slow, or what breaks if you change a type. Codebase archaeology that used to take a day of grepping takes minutes.

Paste the stack trace, the weird screenshot, the failing request. It reproduces the bug, finds the cause, writes the fix and runs the tests. My favourite: it traced washed-out video colors to a missing color-space tag in an encode step I didn't know I had.

Wire hooks so your tests and linter run automatically after every edit it makes. The agent catches its own regressions before you ever see them, which is what makes leaving it unsupervised sane.

A review pass over your diff before it ships catches the off-by-one, the unhandled error path, the API you misused. A senior second pair of eyes on demand, including on the code it wrote itself.

Your most repeated multi-step chore becomes a custom skill: one command runs the whole chain. The biggest compounding habit on this list, because every workflow you wrap stays fast forever.

A tiny VPS with a queue and a cron loop runs your scheduled work: reports, syncs, backups, publishes. Mine happens to post content on a schedule; yours could ship nightly builds. Same architecture.

Anything you do by clicking a web console can be a script against the official API: publishing, provisioning, pulling stats, rotating keys. Claude reads the docs and writes the client so you never click through five screens again.

Bridge your pipeline to Telegram or Slack: outputs come to your phone for approval, a reply triggers the next step. I have reviewed and shipped production work from the gym.

ffmpeg, whisper transcription, scene detection, batch encodes: Claude orchestrates all of it from a description. It just so happens mine cuts talking-head takes into captioned reels; the same pattern processes any audio or video at scale.

Describe the exact view of your system you wish existed: your queue, your metrics, your logs, one tab. Custom internal tools used to need a team and a sprint; now they're an afternoon of conversation.

Raw numbers lie when samples differ in age or size. Have Claude pull from the APIs and normalize: same-age comparisons, per-unit rates, honest baselines. Mine scores every post at exactly 72 hours old for this reason.

Invoices, usage reports and reconciliation drafted from actual API numbers instead of a hand-maintained spreadsheet. Month-end becomes a sanity check, not a session.

Point Cowork at the mess: it files screenshots, renames assets into a scheme you can grep, dedupes downloads, and drafts replies to the boring half of your inbox while you build.

Once a week, put your numbers in front of it and make two decisions: kill what's fading, double what's working. Ten minutes, and it's a standing meeting with an analyst who read everything and forgets nothing.

Describe the site or app in plain English; it scaffolds, builds, deploys and iterates. Push to git, live in a minute. This site, this page included, exists that way.

Calculators, converters, checklists, tiny utilities for your team or your audience now cost a conversation to build. Small tools that save someone ten minutes are the most shareable software you can make.

Keep a small wiki Claude reads at the start of every session: current priorities, active projects, decisions already made. Every conversation starts warm instead of from zero. CLAUDE.md is the same trick at repo level.

Voice-note a messy brain dump after a call or a debugging session and have Claude file it into the right pages, update priorities, and archive the raw note. Thinking stays fast, records stay clean.

READMEs, runbooks and architecture notes generated from the code that actually exists, then kept current as it changes. The docs you were never going to write are suddenly the docs you have.