Watch the video with chapters & transcript →
AI TOOLS
In short: Claude Code works better when you start in planning mode and follow a four-step workflow instead of typing a one-line request. Plan first, let Claude interview you with the AskUserQuestion tool, track progress in a process.md file, and test each feature as it is built. This saves time and tokens because the app is only as good as its plan.

The output of Claude Code is defined by your input, and your input lives in the development plan. Prompted with a one-liner, the tool will happily build an application you are only partly happy with; prompted through planning mode, it builds against an agreed specification, as this demo of a "celebrity friend" web app shows.
Planning mode, activated by typing a forward slash and choosing plan, makes Claude Code write a development plan before touching any code. Your application is going to be as good as your plan is, so approving the plan first saves time and tokens compared with iterating on finished code. The demo's first prompt: create a product specification for the app, use the AskUserQuestion tool to clarify requirements, output a PRD with the full feature list.
The AskUserQuestion tool, described in the Claude Code documentation's best-practices section, makes Claude interview you instead of assuming your requirements. In the demo it asked about the product's core concept, target audience, and intended business model, surfacing things you might not have considered yet: technical implementation, UI/UX, edge cases, trade-offs. If you have no answer to a question, ask Claude to propose motivated recommendations and pick one.
Claude Code has a context window, a limit on how much text can be held in one session, and when that limit is reached a new context window starts. Without a file tracking which features are built and which remain, Claude will not know where to continue in the fresh window. Asking Claude to create and maintain a process.md file at plan-approval time gives every future session a reliable to-do list.
| Workflow step | What it prevents |
|---|---|
| Enable planning mode before coding | Building an app you are unhappy with, then burning tokens on rework |
| Prompt with the AskUserQuestion tool | Claude guessing at audience, business model, and edge cases |
| Create a process.md file | Losing track of progress when the context window resets |
| Test every feature right after building it | Discovering broken features only at the end of the build |
Ask Claude Code to write a test for every single feature it creates and to run that test immediately after the feature is built. In the demo, one message approved the plan, requested the process.md file, and required per-feature tests, and Claude confirmed with "ready to execute".
Ani Björkström, a Stockholm-based tech consultant, summarizes the workflow in four rules: plan first, let Claude ask the questions, track progress in a markdown file, and test each feature as it lands.
Type a forward slash in the Claude Code prompt and choose plan. Claude then drafts a development plan and asks "Would you like to proceed?" before writing code.
No. The demo prompts were drafted in the Claude chatbot and pasted into Claude Code; ChatGPT or any other LLM works the same way.
When the context window fills up and resets, Claude Code loses its picture of which features exist. The process file is what lets it resume exactly where it stopped.
Read the full video transcript, with chapters, on the video page →
Want this working inside your finance team?