Home/Insights/AI Tools

You've Been Using AI the Hard Way (Use This Instead)

By Ani BjörkströmPublished 25 February 2026Reviewed 20 September 20264 min video + articleAI Tools

You've Been Using AI the Hard Way (Use This Instead)
▶ Watch the full 4-minute tutorial · free on YouTube

AI TOOLS

Are You Using Claude Code the Hard Way? Five Best Practices That Fix It

In short: Anthropic's newly published Claude Code best-practices documentation surfaced at least five mistakes even a working consultant was making daily.

Key takeaways

  • Anthropic's newly published Claude Code best-practices documentation surfaced at least five mistakes even a working consultant was making daily.
  • Pre-approving file edits, git commands, and test runs removes the constant permission prompts without resorting to the riskier skip-permissions mode.
  • Subagents run in their own context windows, so delegating small tasks keeps the main agent's limited context usable for longer.

Claude Code ships with safe defaults, and most people never change them — which means they work slower than necessary. After reading Anthropic's new best-practices documentation end to end, Ani Björkström, a tech consultant based in Stockholm, found at least five things she had been doing wrong and distilled them into these five fixes.

How do you stop Claude Code from asking permission constantly?

Either bypass all prompts with the dangerously-skip-permissions option, or — the safer route — explicitly allow the actions you trust: edits to files, git commands, and running tests. By default Claude asks before writing files, committing to git, running bash commands, or using MCP tools — deliberate for safety, but it slows you down massively across a session.

What are skills and subagents actually for?

Skills extend Claude's knowledge with information specific to your project; subagents offload tasks into separate context windows. A skill is a named file holding domain knowledge — for example API conventions like "kebab-case for URL paths, camelCase for JSON properties, always include pagination for list endpoints" — applied automatically when relevant or invoked with a forward slash and the skill name.

Subagents matter because the main agent's context window is limited. A subagent runs with its own context and tools — the documentation's example defines a "security reviewer" on the Opus model, instructed as a senior security engineer to check injection vulnerabilities, authentication and authorization flows, secrets in code, and insecure data handling — so delegated tasks stop cluttering the main conversation.

Why should you let Claude interview you before building?

Because Claude asks about things you have not considered yet: technical implementation, UI/UX, edge cases, and trade-offs. Instead of writing a long specification, start with a minimal prompt — "I want to build a LinkedIn scraping tool, interview me in detail using the AskUserQuestion tool" — and answer its questions.

The interview often surfaces features you had not considered, forcing real product planning before any code exists. That is cheaper than discovering the missing requirement after the build.

What do you do when Claude Code gets stuck or goes wrong?

Press ESC to stop Claude mid-action while keeping your context — useful when you return and find it has been grinding on the same task for hours. Dedicated controls then let you rewind, undo, or reset.

SituationAction in Claude Code
Claude is stuck mid-taskPress ESC once to stop the action and keep your context
You want to go a few steps backPress the ESC key twice to rewind
Claude made a change you dislikeRun /undo to reverse the last change
Starting an unrelated taskRun /clear to reset the context window

FAQ

Is it safe to skip all permission prompts?

The skip-permissions flag is explicitly labeled dangerous, so the better default is targeted allowances: file edits, git commands, and test runs. You keep most of the speed while Claude still asks before anything else.

When should I use a subagent instead of the main agent?

For small, well-defined tasks like a security review, since each subagent runs in its own context window with its own tools and model. That keeps the main agent's limited context available longer.

What is the fastest way to recover from a bad change?

Use /undo to reverse the last change, or press ESC twice to rewind a few steps. For a fresh start between tasks, /clear resets the context window entirely.

Full transcript of the video (718 words, 6 sections)

Chapters: 0:00 Intro and Goals · 0:26 Permissions and Safety · 1:00 Project Skills Setup · 1:38 Subagents for Delegation · 2:36 Let Claude Interview You · 3:28 Stop Rewind and Clear · 4:07 Wrap Up and CTA

0:00 Hey, Autocode just published documentation about the best practices in Cloud Code. I read the documentation so you don't have to and I noticed at least five things that I was doing wrong. My name is Annie. I'm a tech consultant based in Stockholm and in this tutorial I will show you step-by-step how to avoid most common five mistakes in Cloud Code. Let's get started. [music] Best practice number one. Cloud will always ask you about permissions for writing to files, for doing Git commits, for running bash commands, using MCP tools and this is how Cloud is developed for safety reason. However, this is slowing you down massively. So either you can use dangerously skip permissions that will bypass all the permissions and it will be quite quick or if you don't want to take any dangerous action, you can be like me and you can ask Cloud Code to allow edits to different files, allow Git commands and allow running tests.

0:59 Best practice number two. Create skills for domain knowledges. So in here you can see the example where you are creating the name of the skill [music] and then the description and then you are giving the skill a domain knowledge. For example, in this case it's API conventions, use kebab case for URL path, use camel case for JSON properties, always include pagination for list endpoints. Skills [music] extends Cloud knowledge with information specific to your project. Cloud applies them automatically when relevant or you can invoke them directly with forward slash and skill name. Best practices number three. Create sub agents. The sub agent are running commands on their own context window using their own tools and those are amazing.

1:46 If you don't want to clutter your main agent with sub agent and as you know your main agent has limited [music] context window. So if you delegate some small tasks to sub-agent, then your context window for main agent will be used for longer time. So, in here you can see an example. You will define the name, which is security reviewer, then description, review code for security vulnerabilities, then we have tools, and the model is Optus. Then, you are basically creating what agent needs to do. In this case, you are saying you are a senior security engineer, review code for injection vulnerabilities for authentication and authorization flows, for secrets or credentials in code, and for insecure data handling. As soon as you run this in your context window, the agent will be created.

2:35 Best practice number four is about letting Cloud to interview you. Start with minimal prompt and ask Cloud to interview you using ask user question tool. Cloud asks about things you might not have considered yet, including technical implementation, UI/UX, edge cases, or trade-offs. Use this example command, [music] where you say, "I want to build, for example, LinkedIn scraping tool, then interview me in detail using user question tool." Ask about technical implementation and so on. So, here you will identify what it is that you want Cloud to ask you about. And at some cases, it can be different features that you haven't even thought of, and Cloud will still ask you about those features, which will initiate your thinking and planning product actually [music] want to have before even starting to code it.

3:28 Sometimes Cloud can be stuck with the same task for couple of hours. This can normally happen when you leave your computer and come back, and Cloud is still working on the same task. So, in this case, you can just use ESC key, and this [music] will stop Cloud mid-action, keeping your context. And if you want to go a couple of steps back and rewind, [music] then you can click this button twice. And if Cloud updated something that you don't [music] like, you can use undo.that command. That will undo the last chain. And in between different tasks, if you don't want to keep the current context window, you can always use forward slash clear command. Those were five best practices that I wanted share with you. If you want to get the full list, just like this video and comment full list and I will share it with you.

4:16 Have a good day and bye.

Want this working inside your finance team?

Ani Björkström

Ani Björkström — founder of QvantX Sweden AB, a Stockholm consultancy building AI solutions for banks, asset managers and finance teams. Anthropic partner. Every article starts from a real client build, minus the confidential parts. LinkedIn →