The Claude Features Almost Nobody Turns On, Part 2: Let It Act, Automate, and Scale
Part 2 of the guide to getting more from Claude: the tools that leave the chat window. Reading your files, acting in your browser, running on a schedule, installable skills, CLAUDE.md, Claude Code, visual work, and prompt caching for up to 90% cheaper API calls.
Part 1 stayed inside the chat window and made Claude remember you, build real things, reason on the hard calls, and disagree with you when you needed it. Everything there still ends with you reading a reply and doing the work. This part is about closing that last gap, where Claude stops handing you answers and starts doing the work on your machine, on your schedule, at a cost that makes building on it actually viable.
“An assistant that answers you is useful. An assistant that acts for you is a different category of thing.
”
If you have not set up the basics from Part 1, start there, because a Claude that already knows your context makes everything below sharper. I have grouped these by what you are trying to get Claude to do, not by where they sit in a menu.
Let it touch your real work
The fastest jump in usefulness comes from letting Claude out of the sandbox and into the place your work actually lives.
For most people that place is files on your computer, and the desktop app is what reaches them. Instead of pasting a document into a chat box, it reads your real files, edits them in place, creates new ones, and organizes folders. The shift is from describing your work to Claude to having it work alongside you in the same material, which removes a hundred tiny copy-paste round trips a week.
The other place your work lives is the browser, and the Chrome extension covers that. It can see your active tab and act on it: read the page, click, fill forms, and page through results while you do something else. It shines on the repetitive web chores you would never automate properly but quietly lose hours to.
The kind of grind to hand it
I am on this job listings page. Go through every visible listing and extract the title, company, salary range if shown, and the top three requirements. Build a table sorted by salary, highest first. If there are more pages, click through and keep going until you have covered them all.
Make it run without you
Acting on demand is good. Acting without you in the loop is where it starts to feel like leverage rather than a tool.
Scheduled Tasks are the headline here, and they are the feature I wish more people knew about. You define a job once and Claude runs it on its own, at the cadence you pick, and drops the output where you want it. The mental model flips from “open Claude and ask” to “wake up to the finished result.”
Search the last 24 hours of AI news, pick the five stories that matter to builders, write a two-sentence summary and a “why it matters” for each, and save it to /briefs/brief-[date].md. Direct and analytical. Readable in three minutes.
A scheduled task gets sharper when it knows how to do a specialized job, and that is what Skills are for. A skill is a pre-built instruction set that hands Claude a specific capability, building a slide deck, handling PDFs, running a particular workflow, without you re-explaining it each time. They behave like apps for your assistant: the base tool works fine, and the right ones installed make it do far more, reached for automatically when a task fits.
If you build, three that change the economics
The remaining tools are aimed at people who write code or ship on the API, and they are the ones I lean on daily.
Claude Code is the big one. People know you can get code snippets from a chat. Fewer know there is a version that lives inside your development environment, reads your actual codebase, writes changes, runs the tests, reads the failures, and fixes them in a loop until the task passes. It plugs into VS Code and JetBrains and can run inside GitHub Actions to review or open pull requests on its own. This site is maintained with it.
The way you keep Claude Code on the rails is CLAUDE.md, and if you build anything in a repo this is quietly the highest-leverage item across both parts. It is a plain text file in your project that Claude reads at the start of every session, automatically, no prompting. Your conventions, your terminology, your voice rules, your folder layout. Written once, obeyed forever, by every session in that project.
A slice of a real one, for a writing project
# Project: AI newsletter
## Writing rules
- Short paragraphs, three sentences max. Prose, not bullets, in editorial copy.
- No em dashes. Restructure the sentence instead.
- Numbers beat adjectives: “saves 3 hours”, not “saves significant time”.
- Never use: delve, groundbreaking, leverage (as a verb), utilize.
## Content rules
- Assume the reader knows what an LLM is. Lead with the most surprising thing.
- Every piece needs a concrete “so what”: what to do or think differently.
## Structure
- Drafts in /drafts, published in /published as YYYY-MM-DD-title.md
The third is prompt caching, which changes your unit economics if you build on the API. When your calls share a big repeated block, a long system prompt, a reference document, a codebase, you normally pay to reprocess those identical tokens every single time. Caching stores that block so later calls reuse it.
Cost reduction on cached input tokens, on a cache hit
How long the cache lives, the timer resetting on each use
Common latency improvement on long, cache-heavy prompts
You mark the block you want cached, and every call inside the window reuses it. It works for system prompts, large documents, and tool definitions. Running a production workload with a big stable context and no caching means paying full price for the same tokens on repeat. Check the current rates in the Anthropic API docs before wiring it in, since the specifics change.
One more, slightly off to the side of the rest: for visual work rather than code or text, Claude Design turns a description into one-pagers, decks, prototypes, and landing-page layouts you can export. For non-designers it compresses a long session in a design tool into a short conversation.
Now stack them into a pipeline
The feature lists stop at the individual tools. The compounding, again, is in chaining them, and this is the part worth stealing. Picture a single standing pipeline: a Scheduled Task runs each morning, calls a Skill that knows how to research and format a brief, writes the result as a file the desktop app files into the right folder, and on the build side a repo governed by CLAUDE.md and worked by Claude Code turns the best of those briefs into shipped changes. No single piece is exotic. Wired together, one person runs an operation that used to need a small team.
“Each of these is a feature. Chained together they are an operating system for a one-person workflow, which is the thing almost nobody sets up.
”
So do not read this and change nothing, which is the real failure mode. Pick the two that match how you already work and turn them on today. If you live in files and the browser, that is the desktop app and the Chrome extension. If you want leverage while you sleep, a Scheduled Task and a Skill. If you build, a CLAUDE.md, Claude Code, and caching on any serious API workload. The people who get superhuman output from Claude did not find a better model. They configured the one everyone has to know them, act for them, and keep working when they are not watching.
