Give Claude a finish line

Give Claude a finish line

Give Claude a finish line

Read time: 3 minutes

You ask Claude for a client reply.
It comes back close.
Then you fix it by hand.

The fix: write what "done" looks like before Claude starts.

Anthropic's prompting guide calls Claude:

"a brilliant but new employee who lacks context on your norms and workflows."

It also says:

"Be specific about the desired output format and constraints."

A new employee also needs to know when the job is finished.

That is a finish line.​
A short list of things you can see in the draft.
Each one is a yes or a no.

One client question

Illustration: the client and drafts are made up to show the method.

Maria emails:

"Should I raise my prices in January?"

1. Write the finish line

  • The first line answers yes or no.
  • It names the one number she should check.
  • It gives one next step for this week.
  • It is under 120 words.

Every item is something you can point to.​
"Sounds professional" is not on the list.
You cannot point to it.

2. Put it in the prompt

Give Claude Maria's situation first.​
Then add the four items:

It is done when:

+ The context tells Claude who Maria is.
+ The list tells Claude when to stop.

3. Check the draft in a fresh chat

Paste the list and the draft.​
Ask:

For each item, yes or no. For every yes, quote the line that proves it.
The first draft opened with "great question."
​Item 1: No.
The other three: yes.

Now you know exactly which one failed.

↳ Give the fresh chat only the list and the draft.
↳ It only sees the list and the words.

4. Fix only what failed

Do not ask for a new draft. Ask for one change.

Rewrite only the first line so it answers yes or no. Keep everything else.
New first line: "Hi Maria, yes, raise them in January."

Check again.

Four yeses. Send it.

Use the same loop on any draft that matters.

One limit: the checker can be wrong, so read the quotes yourself.

Keep the answer settled

Anthropic notes Claude Opus 5.5 "sometimes goes back over an earlier answer" in multi-turn chat.

For that, Anthropic wrote two sentences to add at the end of a chat system prompt.

They tell Claude to treat earlier answers as done unless you ask about them or point out a problem.
​The goal: Claude stays on your follow-up instead of revisiting the answer you approved.

Use them once a reply passes to ask Claude to treat it as settled.

I put them in a free guide, ready to copy, with where to paste them.

Get the 2 sentences for Claude's instructions:

  1. Go to ericmelillo.com/claude-instructions-2-sentences​
  2. Enter your email
  3. Confirm, and the guide arrives

​Read the expanded edition →​


PS - Whenever you’re ready, here are 3 ways I can help you:

  1. ​Free Resources - Get swipe files, templates & AI prompts to accelerate your growth.
  2. ​Online Courses – Learn step-by-step from proven frameworks to build and scale your creator business.
  3. ​Book 1:1 Strategy Call - Being stuck is a choice. In 30 minutes, we’ll map your next 3 moves to get you unstuck. No cost, no strings. Creators who win don’t wait.

Have questions? Hit reply to this email, and we'll help out!

AI for Leverage

You are already good at the work. The output does not show it yet. Every week I hand you one AI system that gives hours back, already running in my own business, with the numbers attached.