Feed the Beast: AI Eats Software
I was eating pho with a friend in Williamsburg.
He told me his CEO - who’d never written a line of code - was running their company from an AI code editor.
I almost fell out of my chair.
OF COURSE. WHY HAD I NOT THOUGHT OF THAT.
I’ve since gotten rid of almost all of my productivity tools.
ChatGPT, Notion, Todoist, Google Keep, Airtable, Perplexity. All gone.
At first it felt like losing a limb.
I’d been using some of these for almost a decade.
But as the weeks went by, the eerie feeling went away.
Now, my work centers around text files.
The Old Way: Conform to Software
Software interfaces force their way of thinking on us.
Notion wants hierarchy.
A CRM wants structured fields.
ChatGPT has chat threads and projects.
All require clicking, learning the interface, and copy-pasting between tools.
But we now have a technology that can adapt to us and our data.
The New Way: Feed the Beast
“I use AI” -> “AI uses my data.”
Claude Code is the center of how I organize my working life.
An AI that can read and write files on my computer.
My role is to ensure it has the data it needs.
To-dos, meeting notes, roadmaps, specs, philosophy, emails, essays, and more.
The more I give, the more AI can help.
Example 1: Who should I be asking for help?
People overestimate AI’s value with writing, and underestimate its value for reading.
People search is one of my favorite use cases for Claude Code.
I made a custom slash command that searches my network.
Step 1: I type “/network” in the command line
Claude then:
- Reads my to-dos and roadmap
- Predicts what I need help with
- Searches my network for matches
- Suggests who I should reach out to
I then get a list of 10-20 top picks, how they overlap, and their contact info.
The New No Code
I was working with a founder friend at Bryant Park.
I demo’d the /network command, and he couldn’t believe it.
He pointed out what I’d taken for granted: this required (almost) no code.
I dumped text into a file. Wrote a prompt. Spent 15 minutes vibe coding a Python script.
Building an application to do this would be a bear: design the interface, set up a database, connect to APIs, fix bugs, integrate data sources, set up hosting, etc.
Example 2: Weekly Review
Before Claude code I hated doing my weekly reviews.
80% of the time I’d skip them, despite the clarity they brought.
THE OLD WAY
Bounce between 5 software tools to find the data I need.
Write in a Notion Doc.
THE NEW WAY
I run “/weekly” from Claude Code.
- Claude looks at every file change since last week
- Claude evaluates the state of projects, tasks, and the roadmap
- We have a conversation to dig deeper, and make decisions
- Claude generates a document summarizing the week and plan we agreed on
Running /weekly
Claude spots things I’m missing, work I’m procrastinating on, and creates a space to dump everything on my mind.
It then updates files in preparation for the week ahead.
Outputting a Review:
Weekly reviews are now something I look forward to.
Example 3: The Overnight Assistant
Common routine: wake up, check email, Slack, calendar, CRM, Linear, product metrics, Github PRs, etc.
Then get distracted and overwhelmed.
With Claude Code running via GitHub Actions every morning before I wake up, I don’t need to be there to push buttons and sift through data.
MCPs enable hooking into external applications to get the data Claude needs.
While I sleep, Claude can:
- Prepare meeting briefs (past conversations + company research + email history)
- Sift through product metrics and support tickets that may indicate problems
- Look at all inbound data, and help prioritize tasks for the day
When Claude is done synthesizing, it creates a pull request with its findings and suggests next steps.
When I wake up, I review the changes.
Then act on what I want, ignore what I don’t.
Failure Modes
1: Early Days, Rough Edges
“AI Teammates” are not here yet.
LLMs are imperfect by nature.
AI infrastructure is still nascent and often painful to set up.
MCPs are slow, unreliable, and create security vulnerabilities.
My (current) solution: not allowing Claude Code to take external facing actions (e.g. sending an email).
And expecting constant screw ups.
2: Indifferent Amplifier
With AI, you can run yourself off a cliff (or into the promised land) at record speed.
No need for editors, co-founders, therapists, or employees.
If you’re creating something nobody wants, AI will accelerate that.
If you are moving in the right direction, AI will accelerate that.
3: AI Holes
If I asked a human “Is this email draft any good” for 45 minutes straight, they’d throw my computer at the wall.
AI will oblige without complaint.
The solution isn’t asking for “honest feedback” - it’s recognizing when you’re in an AI hole and pulling out.
Getting Started
- Install Claude Code - starts at $20/month
- Create a local project folder - Cursor and/or Obsidian are great for editing files
- Dump everything into folder - Copy docs, notes, todos into .md files
- Initialize Claude from Terminal - navigate to the folder, type “claude” into terminal, hit enter
- Sample first command: “Read all these files and explain what I’m working on”
- Expand from there - Ask Claude to organize, write CLAUDE.md files, create slash commands, Git / Obsidian Sync, schedule tasks, etc.
- Hook in external data via MCP - Slack, HubSpot, Stripe, etc. Docs
If you’d like to duplicate my commands and file structure, see this Github Repo
The Agents are Coming
I inspired a CEO (of a 40-person startup) to run his operations on text files.
His entire team is using Claude Code.
He was surprised by how willing they were to transition.
Customer feedback, sales calls, support tickets, user research, bug reports - it all gets thrown into the pot.
And AI handles how to organize and connect the data.
His comment: “We are enabling AI agents that will one day run the company”
Autonomous Agents
“I use software” → “I use AI” → “AI uses my data”
For most people, they are still the catalyst.
They wake up, search for files, write a prompt for AI, copy-paste data, watch AI work, provide feedback, copy-paste the result.
But that’s changing.
One measure of AI progress: how long can AI do useful work without supervision?
Since Claude Code:
- has all of my data
- can search the internet
- knows my to-dos, and goals
- can plan and find files
It can get started on my tasks without me.
Before I wake up in the morning.
But it’s not quite there yet.
The beast is hungry. Feed it well.
Dive Deeper
Learn about New York AI
Subscribe for updates