I build henny.info, the shop, the apps — and this newsletter — with a tool called Claude Code. It’s an AI assistant that lives in the terminal: it reads your code, edits files, runs commands. The interesting part isn’t that it writes code. The interesting part is what you no longer have to do yourself.

Here are ten tasks I’ve handed off. From “helps in the everyday” to “runs while I sleep.” None of it is magic — and none of it takes the decision away from you. They take away the repetition.

1. Get up to speed on an unfamiliar project

The first thing I use it for isn’t writing at all, it’s reading: “Explain this project to me.” It reads through the code, finds the parts that matter and answers questions — instead of me clicking through unfamiliar folders for hours. For old code of your own that you no longer understand, it works just as well.

2. Save recurring workflows as a “skill”

The things I do the same way over and over — deploy, start a blog post, review the code — I write down once as a skill and then call with /deploy. A skill isn’t a program; it’s a saved set of instructions that it reads and follows. (A small clarification, because it’s often told wrong: custom slash commands aren’t a feature — anything starting with / is a skill.)

3. Check automatically on every change (hooks)

A hook is a command of your own that always runs on a particular event — for example the code formatter after every file edit. The difference from a skill matters: a hook fires unconditionally, not “if the AI remembers to.” So you can’t forget the formatting or the test anymore — it happens on its own, every time.

4. Connect your tools (MCP)

Through MCP servers it gets access to services outside your machine: GitHub, Slack, a database, the browser. For most common services there are ready-made connectors — you don’t build anything, you connect. After that it can read a GitHub issue, open a pull request or post a message to a Slack channel, straight from the terminal.

5. Run several helpers in parallel (subagents)

For research or large searches it sends out subagents — each with its own head, each returning only the summary. That keeps the main conversation clean instead of stuffing it with a hundred files. No background wizardry: they run off and report back when done.

6. Build Claude into scripts (headless)

With claude -p "..." it becomes a building block on the command line — you can pipe data through it like any terminal tool:

git diff | claude -p "find typos in this diff"

That’s the shift from “I’m chatting with an AI” to “the AI is a step in my workflow.” It runs the task, prints the result and exits.

7. Review pull requests automatically (GitHub Actions)

In CI — GitHub Actions, in my case — it looks over every pull request: for bugs, for security holes, for careless mistakes. It writes its comment straight onto the PR. You can also address it there with @claude and ask for something. So every change gets a second pair of eyes before it’s merged.

8. Run on a schedule or overnight (routines)

Routines run on a schedule in the cloud — even when your own machine is off. Summarize the open pull requests every day, send a short report, watch something and let you know. (For quick repetition inside an open session there’s /loop — but that’s for the moment, not for standing operation.)

9. Hold on to project knowledge (CLAUDE.md + memory)

A file called CLAUDE.md holds the project’s rules — “use pnpm, not npm”, “tests before every commit”, “the API endpoints live here” — and they apply automatically in every session. Plus a memory it uses to note down what it has learned. You explain things once, not every time.

10. Allow trusted commands (permissions)

In a settings.json you decide what it may do without asking — git status, running the tests, looking at the state. Everything else it keeps asking about. So it works smoothly on the harmless things without you giving up control on the risky ones. (If you want to hard-forbid something, you reach for a hook again — it can block an action, whereas a permission only allows one.)

The thread running through all ten: you say one thing once, done right — as a skill, a hook, a rule — and after that it happens on its own. That’s exactly why a single person today can build and run a whole shop that used to take a team. This site, the shop and the newsletter are the running proof.

I’m still learning this myself every day, and the tools change fast — the feature status here is July 2026, and tomorrow something may have a different name. If one point interests you more, write to me; I’ll turn it into its own post.

Sources

The feature names and examples are checked against the official Claude Code documentation (as of July 2026): Skills, Hooks, MCP, Subagents, Headless mode, GitHub Actions, Routines, Memory, Permissions.

Questions, or found a mistake? Write to me.