This is my exact AI setup to write better prompts:

(steal it in 2 minutes):

____

1. Create a Project (stop using blank chats).

→ Go to Claude → Click "Projects" → New Project.

→ Add custom instructions: tone, audience & goal.

→ Upload 1-3 reference files (download as .md file).

→ Use this set up guide: ruben.substack.com/p/i-…

____

2. Turn on Extended Thinking (before you prompt).

→ Click the "Thinking" toggle below your chat input.

→ AI stops pattern-matching. Starts reasoning.

→ Same prompt, completely different depth.

→ Yes, it's slower. But also better. Don't be a baby.

____

3. Turn on Search (if you need accuracy).

→ Click the "Search" toggle next to Thinking.

→ AI stops guessing. Starts citing real sources.

→ This is the antidote to hallucination.

→ Al with eyes lies far less than one running blind.

____

4. Upload a reference (stop describing yourself).

→ Find an example of exactly what you want.

→ Screenshot it or download as markdown.

→ Upload it and type: "Match this tone & structure."

→ Copy-paste my prompt: ruben.substack.com/p/i-…

____

5. Use AI as a critic (not a creator).

→ Write a rough draft yourself first.

→ Prompt: "What's weak about this? Be brutal."

→ AI explains well and executes generic.

→ Copy this critic prompt: ruben.substack.com/p/yo…

____

6. Define success (not steps).

→ Don't tell AI how to do it. Tell AI what is "done".

→ Add: "Who is this for? What does it look like?"

→ Add: "What should it NOT sound like?"

→ Outcomes > process. Always.

____

7. Specify constraints.

→ Tell AI what to avoid, not just what to include.

→ Add: "No fluff. No corporate tone. Under 150w"

→ Add: "Don't include X, Y, Z."

→ Boundaries > rules.

____

8. Give examples.

→ Paste a good example directly into the chat.

→ Type: "This is the tone I want. Match it."

→ Show "too shallow" vs "just right."

→ It learns from what you show.

____

9. Reset after 15 messages.

→ Context gets bloated. It starts drowning.

→ Start a new chat inside the same Project.

→ Only carry forward what actually matters.

→ Less = more. Always.

____

Now here's how to know you're doing it wrong.

11 signs of a bad prompt:

✦ You start in a blank chat, not a Project.

✦ No thinking mode enabled.

✦ No reference uploaded.

✦ Vague goal like "make it good."

✦ One-shot prompting, no back-and-forth.

✦ You expect AI to fill in the gaps.

✦ You ask AI to create, not critique.

✦ No success criteria defined.

✦ You describe steps, not outcomes.

✦ Bloated context, never reset.

✦ You give too much info (less = more).

____

Most people don't know how to set up AI correctly.

So I wrote the full setup guide (for free).

To access it: ruben.substack.com/p/yo…