Claude Code does what you write, not what you mean. In this part, you learn three ways to close this gap: plan first, say what “done” looks like, and tell Claude Code what it cannot see. Each one is a sentence or two more in your prompt, and it saves you rounds of corrections.
3. Plan first when you are not sure what you want
Not every task needs a plan. Which way is best depends on how well you know what you want.
Quiz. Which way for which task? Match each task with the way that fits it.
| Task | The way |
|---|---|
| The draft of your flow greets every customer with “Dear Sir”. You know which flow it is. | One prompt: what goes wrong, where, and how you will know that it is fixed |
| Your team wants a new feature: your form asks for the order number, your flow uses it, and n8n sends urgent orders to a person. You have not thought it through yet. | Let Claude Code ask you questions and write a plan, then let it build the plan |
| Three small changes that have nothing to do with each other | One prompt for each, and /clear in between |
Why
For the greeting, you know what is wrong and where: one clear prompt is enough. Three changes that have nothing to
do with each other go one by one, with /clear in between: each starts with a clean conversation.
A new feature has questions that only you can answer: which field, what counts as urgent, who gets the mail. Claude Code cannot see your answers. Without the interview, Claude Code answers the questions itself, and you find out at the end. When Claude Code asks you first, you catch a wrong idea in the plan, where it is quick to fix, and not later in the work.
Your turn: plan one small thing. Here you only plan it; you do not build it. The thing you want: a short webpage for a new member of your team, which says what your project does, which flow does what, and how to try it out. The plan shall also say how you will know that the webpage is complete.
New here? Plan a webpage about a team of yours
You have no team project of the course. So plan a webpage about a team or a group that you are in, at work or
elsewhere. Start Claude Code in your folder ~/projects/practice, and paste this prompt in place of the one in 1:
I want a short webpage for a new member of my team: what we do, who does what, and how to start. I have not thought it through yet. Interview me until you could write a plan, then write the plan to plan.md. End the plan with how we will know that the webpage is complete. Do not build anything yet.
Claude Code writes plan.md into your folder. To open it in 3, paste this line on a Mac:
open -e ~/projects/practice/plan.md
On Windows:
cd ~/projects/practice && notepad.exe plan.md
-
Start Claude Code and paste this prompt. The sparkles before it stand for Claude Code: they are not part of the prompt.
I want a short webpage for a new member of my team: what our project does, which flow does what, and how to try it out. I have not thought it through yet. Interview me until you could write a plan, then write the plan to plan.md. End the plan with how we will know that the webpage is complete. Do not build anything yet. -
Answer its questions, until Claude Code has written
plan.mdinto your course folder. Count the questions that you had not thought about yourself. Does Claude Code ask whether it should build now? Say no: here you only plan. -
Read
plan.md. Open a second terminal tab (how) and paste, on a Mac:open -e ~/kingo-pod/claude/plan.mdOn Windows:
cd ~/kingo-pod/claude && notepad.exe plan.mdDoes its end say how you will know that the webpage is complete? Is something missing or wrong? Tell Claude Code, and let it change
plan.md.
Claude Code asks to switch to plan mode?
Plan mode is for exploring: Claude Code searches and reads, but it changes nothing. It may still ask for your OK to search or to read something.
You do not need plan mode here: the prompt already asks for a plan, in a file. So you can say no. Did you say yes?
Then, when its plan is ready, Claude Code asks whether it should go ahead and build. Choose Yes, manually approve
edits: then Claude Code asks you before each change. Say yes when it writes plan.md, and no to everything else:
here you only plan. No plan.md yet? Ask: “Write the plan to plan.md.”
Done: you have a plan, and maybe some questions that you had not thought about.
Next time. When you are not sure what you want, start with this interview. To build a plan, start fresh: type
/clear, so that Claude Code begins with the plan and nothing else, and then ask it to build what plan.md says. You
do that in step 10.
- InterviewClaude Code asks, you answer
- plan.mdthe answers, in a file
- /clearthe conversation goes, the file stays
- Buildfrom plan.md, until it is done
/clear empties the conversation; plan.md stays, because it is a file.From Anthropic’s best practices for Claude Code: brainstorm with Claude first; on plan mode: permission modes.
4. Say what “done” looks like
A finish line is one sentence that says how you will know that a task is done.
A real case from our tests. In the last homework, the mail AI assistant drafted a reply to the mail “Is there any Korean home made food?” and did not use the information from an older mail, “Kimchi for Monday”. We asked Claude Code to make a copy of the Langflow flow Draft Reply (Enron) and to improve the search in the mail database in the copied flow. We gave Claude Code no finish line.
We were not lucky. This is what happened:
-
Claude Code set its own finish line:
Run the copy once with the same mail and check that “Kimchi for Monday” now appears in the sources.
The sources are the mails that the search finds for the draft.
-
Claude Code ran the copy. “Kimchi for Monday” was among the sources.
-
The mail AI assistant wrote the new draft. Claude Code read it and told us what it said:
I am afraid I still don’t have any information about Korean home made food.
Quiz. Claude Code reached its own finish line. Is the job done?
- Yes: "Kimchi for Monday" is among the sources, as Claude Code checked
- Not for you: you wanted a reply that uses the kimchi mail
Show the answer
Not for you. You decide what “done” means. Claude Code kept its word: it checked what it said it would check. But its finish line was smaller than yours. With your finish line, Claude Code can check its work against what you want: when the check fails, Claude Code fixes and checks again.
Quiz. Which finish line should we have given Claude Code?
- You are done when "Kimchi for Monday" is among the sources.
- You are done when the search works better.
- You are done when the copy of the flow runs without an error.
- You are done when the draft for my mail "Is there any Korean home made food?" mentions the kimchi from the mail "Kimchi for Monday".
Show the answer
The draft mentions the kimchi. This finish line names one input, your mail, and what should come out: a draft that mentions the kimchi. Claude Code can run the copy and check that by itself. The other three are not enough:
- “Among the sources” is the finish line that Claude Code set itself: the draft still had no kimchi in it.
- “Works better” gives Claude Code nothing to check.
- “Without an error” says that the flow runs, not that the draft is right.
Claude Code’s finish line: "Kimchi for Monday appears in the sources." The check passes, and Claude Code stops. The draft still says that there is no information about Korean food.
Your finish line: "The draft mentions the kimchi." Now the same run fails the check, so Claude Code is not done yet: it fixes and checks again.
Ask for proof. When Claude Code says that it is done, ask: “Show me what you ran and what came out.” “It works” alone tells you little, and reading the output is faster than checking the work yourself.
Done: you can tell Claude Code’s finish line from yours.
Next time. Add two sentences to your next prompt: “You are done when …”, with one input and what should come out, and “Show me what you ran and what came out.”
From Anthropic’s best practices for Claude Code: give Claude a check to run, and ask for evidence.
5. Tell Claude Code what it cannot see
Claude Code can read your files, and in the course also your flows, your workflows and your database. It cannot read your head: what went wrong, where to look, what must stay as it is, and what you want back.
In the last homework, step 12 gave Claude Code this prompt:
My mail AI assistant drafted a reply to my mail “Is there any Korean home made food?” without the mail “Kimchi for Monday”, which is in the database. Find out why.
Look at:
- the last run of WS2.4 - Reply with Approval in n8n: what it sent to the Langflow flow Draft Reply (Enron), and what came back as sources;
- the code of the component Mail Search in that flow.
Do not change anything. Then explain the cause to me in plain words.
This prompt tells Claude Code what it cannot see. Look at its pieces:
Quiz. What does each piece tell Claude Code? Match each piece of the prompt with what it says.
| Piece of the prompt | What it tells Claude Code |
|---|---|
| “drafted a reply to my mail … without the mail ‘Kimchi for Monday’, which is in the database” | What went wrong |
| “the last run of WS2.4 - Reply with Approval in n8n”, “the code of the component Mail Search in that flow” | Where to look, with full names |
| “Do not change anything.” | What must stay as it is |
| “explain the cause to me in plain words” | What you want back |
Why
Each piece is something that Claude Code cannot see by itself. With all four, Claude Code knows what went wrong, where to look, what to leave alone, and what to give you back. The recipe below has the same pieces. Its last piece, the finish line, says what you want back, in a way that Claude Code can check.
Your turn. Make the prompt “Fix the reply.” as good as the one above: in the picture below, add the four pieces, one after the other.
Fix the reply.
- What goes wrong The draft for my mail "Is there any Korean home made food?" says that there is no information about Korean food, but the mail "Kimchi for Monday" is in the database.
- Where to look Look at the Langflow flow Draft Reply (Enron), and at the last run of WS2.4 - Reply with Approval in n8n.
- What must stay Do not change the original flow: work on a copy.
- Your finish line You are done when the draft for that mail mentions the kimchi from the mail "Kimchi for Monday". Show me what you ran and what came out.
A recipe for your own prompts
<What goes wrong, with one example>.
Look at <the full name of the file, or of the flow or workflow and whether it is in Langflow or in n8n>.
Do not change <what must stay as it is>.
You are done when <how you will know that it worked>.
Is there a file that Claude Code should read, for example plan.md from step 3? Type @ and the file’s name in
your prompt: the tool adds the file to your prompt.
Side note: first the smallest version, then the next step
Claude Code reads a request carefully and builds the smallest thing that does the job: it cannot see how much you hoped for. Often that is just right: you check the small version, then you ask for the next step, and so on. When you know from the start that you want more, say so, for example: “be ambitious: also handle mails without a signature”.
Done: you turned “Fix the reply.” into a prompt that tells Claude Code what it cannot see.
Next time. In your next prompt, write what went wrong with one example, and the full name of the file, flow or workflow.
From Anthropic’s best practices for Claude Code: write a specific prompt.
Next: Part C · Stay in control →
If something goes wrong
Find your step.
Step 3
| What you see | What to do |
|---|---|
| Claude Code starts to build the page, not only the plan | Press Esc, and write: “Only plan: write plan.md, and build nothing.” |
After the interview, there is no plan.md | Ask Claude Code: “Write the plan to plan.md.” |