The skill kit for the three build stages with Claude Code

1. Where are you right now?

Do not paste everything at once. Pick the stage you are at, copy its prompt, and only move to the next stage when you finish this one.

• No code written yet? Start with planning.
• Have an organized plan? Move to execution.
• Claude says it is done? Move to mapping and verification.

2. Install the skills once

The seven skills in this guide come from the official Superpowers plugin. Inside Claude Code, in the chat input line, paste the command and press Enter. This is a Claude Code command, not a terminal command.

• When it finishes, an install message appears. Type /plugin and look for Superpowers under Installed.
• Open a new conversation after installing. The skills trigger based on the task, without you needing to remember their names.

/plugin install superpowers@claude-plugins-official

3. Plan before you code

The skills: brainstorming and writing-plans. They help nail down what you are building, who it is for, what must work, and what is out of scope for now.

Before you write any code, brainstorm with me first.

Ask me one question at a time, sharpen the requirements, and present the options with their tradeoffs.

Once we have locked the direction, write an organized work plan with stages, files expected to change, tests, and completion criteria.

Do not start executing before you have shown me the plan.

4. Execute without losing direction

The executing-plans skill works from the plan. systematic-debugging reads the error and looks for the root cause before fixing anything. On large tasks, subagent-driven-development splits the work between subagents and checks each part.

Execute the work plan to the end.

Break it into small tasks, and use subagents only when it genuinely shortens or improves the work.

After each part, run the relevant check. If something fails, read the error and the logs, reproduce the issue, and find the cause before proposing a fix.

If something in the plan does not work, explain what changed and update the plan. Do not silently skip a step, and do not stop just because you hit a snag.

5. Map and verify before you approve

The verification-before-completion skill requires a run and proof before declaring the work done. requesting-code-review checks every change and returns findings to fix. Now Claude has to show what it built, not just write "done".

Map out what you built for me.

Explain the flow, the files, and the key components, and compare the result to the original plan point by point.

Run the system and the relevant tests. For everything you mark as done, show me proof.

Then do a code review on every change, fix the findings, and write explicitly what is still open.

6. A one-minute success check

• There is a plan you can compare the result against.
• Every completed step got a real check, not just a success message.
• The mapping explains where each part lives and how the flow works.
• Gaps and issues are written down explicitly, not hidden behind "done".

7. If Claude still stalls halfway

• Ask it to read the error and the logs before it proposes a cause.
• If the plan is too big, split it into a part you can run and check right now.
• If it changes direction, ask it to update the plan and explain why.
• If it claims to be done without proof, go back to the mapping and verification prompt.

8. The short order

Plan. Execute. Map and verify. Only then approve that it is done.