Five Months In: What Actually Ships and What Gets Abandoned

Every founder I talk to wants the tool list. What's in your stack? What do you use daily? After five months of building production systems with AI agents, I have an answer that nobody likes: most of what I tried got abandoned. Only a few tools survived.

The Survivor Test

My test for whether a tool stays in the workflow is brutally simple. Did I use it this week without thinking about it? If the answer is no, it's gone. Every tool competes for cognitive budget, and cognitive budget is the scarcest resource I have.

Tools that passed: Claude Code, git, ripgrep, a plain text notepad. Tools that failed: every shiny IDE plugin, every AI-powered productivity app, most of the agent frameworks I experimented with in May and June.

What the Survivors Have in Common

Looking at the three tools I actually use, they share traits:

• They compose with other tools via text I/O
• They don't own my data or require cloud sync
• They have sensible defaults so I rarely configure them
• They improve the work directly, not the workflow around the work

# My entire daily stack in four lines
claude-code  # the brain
git          # the memory
rg           # the search
nvim         # the canvas

Why Most Tools Fail the Test

Most tools fail because they add a layer without removing one. A workflow app that promises to manage my Claude Code sessions is a second thing to manage. A project dashboard that summarizes my git activity is a second source of truth competing with git log.

The pattern I keep seeing: if a tool doesn't collapse two things into one, it's adding friction. September is when I finally accepted this and stopped trying to optimize my setup. Now I just use it.

For my broader take on tool selection, see the Claude Code documentation.