10 Claude hacks I wish I knew before

You’ve probably felt this:

You open a new chat. You explain the same thing for the fifth time. The output is fine, not great. Somewhere around message 40, Claude starts forgetting what you told it in message 3.

None of that is Claude being “dumb.” It’s you missing a few habits that fix all of it.

Here are 10.

Found a prompt or process that actually works? Don’t retype it every time.

A Skill is a saved set of instructions Claude can pull up on command, like a reusable recipe. Build it once, and every future chat gets that same quality without you re-explaining the steps.

If you do something more than twice, it should probably be a Skill.

Before Claude writes anything, tell it to ask YOU questions first.

Set this as a standing instruction, not something you ask for every single time: “Before starting any writing task, ask me clarifying questions first.”

Now Claude gathers context before it guesses. Way fewer rewrites.

The longer a chat gets, the worse Claude performs in it. Old instructions get buried. Context gets muddy. Claude starts contradicting itself.

The longer the chat, the dumber the model gets.

Start a new chat per workstream. One thread per project, not one giant chat for your whole life.

Working on something bigger than one sitting? Before you end the chat, ask Claude to write a handoff doc: what’s done, what’s in progress, what’s next.

Drop that into your next chat, and the new session picks up right where you left off. No re-explaining the whole project from scratch.

I go deeper on this here :

The regular chat window is fine for quick stuff. But for anything with multiple steps, files, or tools involved, Cowork and Claude Code just do more.

More bang for your buck, literally. Same subscription, way more horsepower.

Stop pasting API keys directly into the chat.

A .env file is just a plain text file that holds your keys and passwords, separate from your code. Claude can read from it without the key ever showing up in your conversation.

Safer, cleaner, and it’s one less thing to accidentally leak.

If an explanation is getting long and dense, ask Claude to show it instead of just telling it.

Claude can build SVGs, Mermaid diagrams, or interactive HTML visuals right in the chat. A flowchart beats 3 paragraphs of description every time.

Only need the words from a screenshot? Copy the text out and paste that instead of uploading the image.

Images cost way more tokens than text.

Connectors hook Claude up to tools outside its own ecosystem, like your calendar, your docs, your project management tool.

Claude can read your actual world instead of just working off what you paste in.

Claude will tell you it’s “done” the moment the output sounds plausible. That doesn’t mean it’s correct.

Give it a real way to verify: a test to run, a spec to match, a set of criteria to hit. Then tell it to keep going until it actually passes, not until it sounds finished.