One email. No follow-up unless I ship something that fits.
Read firstThis is a method for Claude Code, Anthropic's coding agent. The three steps come from a short video, Claude Code Clearly Explained. The prompt wording below is mine, built from its description. The 40 to 50 percent restart point is the video's rule of thumb, not an Anthropic number. Everything about the question tool and the commands is checked against Anthropic's Claude Code docs (25 Oct 2026).
WHAT YOU NEED
Claude Code, installed and signed in. Anthropic's docs at code.claude.com/docs cover setup.
A project folder. An empty one is fine. Open a terminal in it and run claude.
One line about what you want. Not a spec. The interview builds the spec.
STEP 1
Make Claude interview you
The video's one sentence is: interview me about this project, cover trade offs and edge cases, and write no code until I say go. Here it is as a full prompt. Paste it, then fill in the first line.
paste into claudethe interview prompt
I want to build: [one line about the app].Before you write any code, interview me about this project using your question tool (AskUserQuestion).Cover technical details, trade offs and edge cases. Ask in small rounds, with multiple choice options where you can.Write no code and change no files until I say go.When I say "wrap up", write what we agreed into PLAN.md as the spec.
Then answer the multiple choice questions. Your answers are the spec. When the questions stop surprising you, type wrap up, read PLAN.md, fix anything wrong, and only then type go.
The toolClaude Code's built in AskUserQuestion tool asks you multiple choice questions. You pick an option or type your own through the Other row. In the Claude Code package I checked (version 2.1.197, 25 Oct 2026) one call carries 1 to 4 questions with 2 to 4 options each. Per Anthropic's docs, a question stays open until you answer it, unless you set askUserQuestionTimeout to 60s, 5m or 10m. After that, Claude proceeds on its own judgment.
UntestedI confirmed the tool exists and how it behaves from Anthropic's docs and the Claude Code package. I have not run this exact prompt in a signed in session, so run it once on a throwaway folder before you rely on it. If Claude asks in plain text instead, add: use the AskUserQuestion tool.
STEP 2
Never say "build me an app"
Ask for one feature at a time, then test it before the next. The video's version is features A, B and C. Use the plan template below, then paste this:
paste into claudeone feature at a time
Read PLAN.md. Build feature A only.When it is done, tell me exactly how to test it, and wait.Do not start feature B until I say feature A passes.
When A passes, say so, and ask for B with the same words. After each feature, ask one more thing: check that A still works.
THE ONE PAGE PLAN
Plan template: copy into PLAN.md
Fill the top, then let the interview fill the rest. Keep it to one page. If it grows, the features are too big.
PLAN.mdtemplate
PROJECTOne line: what it is and who it is for.Done means: [what I can do when it works]DECISIONS FROM THE INTERVIEW1. [decision] because [reason]2. [decision] because [reason]Trade offs I accepted: [list]Edge cases to handle: [list]Out of scope for now: [list]FEATURE A: [name]What it does:How I will test it:Status: not started / built / passedFEATURE B: [name]What it does:How I will test it:Status: not started / built / passedFEATURE C: [name]What it does:How I will test it:Status: not started / built / passed
STEP 3
Check how full the session is
Type /context. Anthropic's docs describe it as a view of what is filling the window, with capacity warnings. The video's rule of thumb is that quality drops around 50 percent, so start fresh at 40 to 50 percent. Treat that as a prompt to look, not a law. I found no Anthropic number for it.
command
what it does, per Anthropic's docs
/context
Shows what is filling the context window.
/compact
Summarises the conversation to free space, same session.
/clear
Starts fresh on a new task and keeps project memory.
/resume
Returns to an earlier conversation.
/plan
Switches into plan mode before a large change.
FRESH SESSION CHECKLIST
What to write down before you restart
A fresh session forgets the chat. It does not forget files. So write the state into a file first. Paste this before you run /clear:
paste into claudebefore you restart
Update NOTES.md so a fresh session can continue without this chat. Include:1. What is built and tested, and what is not.2. What we are doing next.3. Decisions made, and why.4. Open bugs and anything I should double check.5. The commands to run and test the project.6. The files that matter most.
1checkRun /context. Is it near 40 to 50 percent, or has Claude started forgetting your rules?
2saveFinish or park the feature. Do not restart in the middle of a half built one.
3writeHave Claude update NOTES.md with the prompt above. Read it.
4clearRun /clear. Or /compact if you only need space, not a clean slate.
5resumeOpen with:Read PLAN.md and NOTES.md, then continue with the next feature.
THE TRAPS
What goes wrong, and the fix
Claude starts coding mid interview
Say: stop, no code until I say go. Then carry on answering. Put that line in the first message every time.
Questions are vague or endless
Say: only ask about things that would change what you build, and stop when you have enough.
No question tool appears
Ask for it by name. In Anthropic's don't ask permission mode the tool is denied, so use a normal mode for the interview.
Feature B breaks feature A
Test after every feature, and ask for a check of the earlier ones before you move on.
The plan keeps growing
Cut features until each fits a test you can run in a few minutes. Move extras to out of scope.
You restart and lose the thread
You skipped NOTES.md. Next time write it first. The video's tip is to restart before the session gets heavy.
That is the whole method
Interview first, one feature at a time, check the window, write notes, restart. Sources: the short video and Anthropic's Claude Code docs.
I'm Raj Singh. I've led ecommerce at two Fortune 100 companies, and at one of them I ran about $8M a year in ads for four years. I've worked with AI since 2019. Now my engineering team and I ship AI products, including debtops.ai. Each guide here is written from a real install, with the output dated.
