How I use Claude to run my startups
The workflows, prompts, and systems I use to make Claude feel less like a chatbot and more like infrastructure.
* * *
👋 Hey, I’m Ian. I write about the shift from software as a tool to software as a worker — AI agents, autonomous workflows, and what happens when companies start putting agents on payroll.
I’m also building Narrareach, a distribution platform for writers and founders. It lets you schedule Substack Notes and articles, cross-post to all social media platforms, publish across Medium and Substack, and see what’s actually working so you can double down on the right ideas.
P.S. If you’re building an agent, hiring AI into a workflow, or trying to grow your writing without manually posting everywhere, reply to this email. I may feature a few interesting examples in future deep dives.
* * *
Most people still use Claude like a search engine.
Open a chat.
Ask a question.
Get an answer.
Close the tab.
Which explains why so many people bounce between thinking Claude is incredible one day and strangely disappointing the next.
The difference usually has less to do with the model itself and more to do with how little context it’s working with.
Claude is unusually good at adapting to the way people think, but most users never give it enough information to do that well. Every new conversation starts from zero. No memory. No understanding of what you do, how you write, what kind of answers you prefer, or what problems you repeatedly run into.
So the output stays generic.
What changed things for me was realizing Claude becomes dramatically more useful once it stops behaving like a chatbot and starts behaving more like a persistent workspace.
Instead of repeatedly asking for isolated outputs, you begin building systems around recurring friction.
Writing.
Research.
Decision-making.
Repurposing.
Analysis.
The people getting disproportionate value from Claude are usually doing one thing differently:
They are not opening random chats.
They are building context.
And once enough context accumulates, the interaction changes completely.
Claude stops sounding like a tool you occasionally use.
It starts feeling closer to infrastructure.
* * *
Stop opening random chats
This is probably the biggest mistake most people make.
A normal Claude chat resets every session. Which means every new conversation starts with:
- no memory
- no understanding of your workflow
- no context about your goals
- no awareness of how you think
So Claude defaults to generic answers.
Projects change this completely.
Once you start using Projects properly, Claude gradually understands:
- how you structure ideas
- what you repeatedly work on
- how detailed you prefer responses
- the kinds of problems you care about
The interaction changes completely.
Claude stops sounding like a tool you occasionally use.
It starts feeling closer to infrastructure.
Prompt: Tell Claude who you are
My name is [your name].
I work as [your role].
My main responsibilities are:
- [responsibility 1]
- [responsibility 2]
Right now my biggest goals are:
- [goal 1]
- [goal 2]
I mainly use Claude for:
- writing
- analysis
- coding
- research
How I prefer responses:
- direct
- concise
- challenge weak assumptions
- avoid corporate language
Things I care about:
- [topics]
Things I dislike:
- generic motivational language
- long disclaimers* * *
Most people ask Claude for answers instead of collaboration
Most people use Claude transactionally.
“What is this?”
“Summarize this.”
“Write this.”
The better workflows are collaborative.
Instead of asking:
“What’s the best strategy?”
Ask:
“Tell me where this strategy breaks.”
The strongest Claude users are usually:
- stress-testing ideas
- reasoning through tradeoffs
- identifying blind spots
- forcing clearer thinking
Not just generating text.
Prompt: Make Claude challenge your thinking
Here is my idea:
[paste idea]
Your job is to attack it.
Find:
- weak assumptions
- missing context
- where this could fail
- what I might be underestimating
- what someone smarter than me would immediately question
Do not soften the critique.* * *
Ask Claude to ask questions first
Most bad Claude responses come from one issue:
Claude started too early.
Get Carson’s stories in your inbox
Join Medium for free to get updates from this writer.
Without enough context, the model fills gaps with assumptions. Then people spend time correcting outputs that could have been accurate from the beginning.
Prompt: Ask questions before starting
Before you answer, ask me the 5 most important questions you need in order to do this well.
Do not start solving the task yet.
First gather the missing context.This works especially well for:
- writing
- product planning
- strategy
- technical architecture
* * *
Claude writes badly when it doesn’t know how you write
A lot of AI writing sounds generic for a simple reason:
People expect Claude to sound like them without showing it how they sound.
If you don’t provide examples, Claude defaults to its own writing patterns:
- polished
- symmetrical
- overly explanatory
- slightly corporate
The strongest use case here is not:
“AI writes for me.”
It’s:
“Claude reduces how much editing I need afterward.”
Prompt: Style cloning
Here are examples of my writing:
[paste sample 1]
[paste sample 2]
[paste sample 3]
Analyze:
- sentence structure
- pacing
- vocabulary
- rhythm
- tone
- paragraph style
After this, whenever I ask you to write something, match these patterns instead of defaulting to generic AI writing.The goal is not imitation.
The goal is reducing friction between your thinking and the final draft.
* * *
Claude becomes much more useful inside repeated workflows
The strongest AI use cases are increasingly attached to recurring friction.
Not novelty.
Repeated friction:
- rewriting the same explanation
- restructuring articles
- formatting workflows
- repetitive analysis
- turning one idea into multiple formats manually
This is also part of the reason I’ve been thinking so much about how writers repurpose content for Substack Notes lately.
Most people already have articles.
The hard part starts afterward:
- turning articles into Notes consistently
- keeping the Notes feed active
- figuring out which Notes actually bring subscribers
- cross-posting without turning it into a second writing session
That repeated workflow friction is where tools like Narrareach become useful.
Not because they magically create ideas, but because they reduce the repetitive work that slowly drains creative energy after publishing.
That workflow increasingly looks like:
Article → Notes → distribution → signals → decisions
Not:
Publish once → disappear → repeat.
* * *
Use Claude to think through problems, not just retrieve information
Claude becomes dramatically more useful once it removes repeated thinking.
Not intelligence.
Repeated thinking.
One of the strongest prompts I’ve used for difficult decisions is this:
Prompt: Stress-test a business idea
I want to stress-test this idea before investing more time into it.
Here is the idea:
[paste idea]
Your job is to:
1. Find weak assumptions
2. Explain why users may not care
3. Explain where competitors could beat this
4. Identify the hardest part operationally
5. Tell me what would need to be true for this to work
Be specific.
Avoid generic startup advice.* * *
The actual point
Claude is not smarter than you.
It does not replace thinking.
It does not replace taste.
What it has is:
- infinite patience
- broad context
- the ability to reason through problems from multiple angles
- and the ability to stay consistent across repeated workflows
Most people still use Claude transactionally.
Open chat.
Ask question.
Get answer.
Leave.
The interesting part starts when it becomes part of how you work.
* * *
