The trap of tool-shaped objects

I started setting up my Obsidian vault on a flight from JFK to SFO. Folders, tags, templates, a clipping pipeline. It was a fun way to spend five hours.

Will Manidis has a name for this: tool-shaped objects. Things that look like the right move. They keep you in motion without moving you forward.

Boredom is uncomfortable, but it's good. It's the discomfort that makes you act. Tool-shaped objects remove that discomfort. You stay busy. You never sit with the idleness that would have pushed you somewhere real.

LLMs are the most sophisticated version of this. Their verbal fluency creates the sensation of thinking, processing, making progress. They confirm every instinct we already have, dressing it up as insight.

"Ask what the number is before making it go up."

What I actually want

I don't want to be validated or have my instincts confirmed. I want a thought partner that will actually push back.

To be opinionated is to have something to stand on. Most LLMs don't. They're optimized for agreement and will meet you wherever you are and stay there.

This is what I've been trying to build against.

Personal agent systems

Jason Li's Pixel at Ramp is a personal agent that appears to have interesting opinions. Pixel writes its own essays: see the Five-X Gap piece. Pixel's argument: code has mechanical verifiability. Design requires human judgment to close the loop. Two individually great screens can make a terrible flow. Composition requires taste.

What makes up an agent?

I've been building my own and thinking through the anatomy:

Personality. Who the agent is, how it speaks, what it pushes back on.

Context. Who you are, how you work, what good looks like for you specifically.

Goals. Not tasks. What "becoming better at this" means over time.

Doubt protocols. Self-questioning for known failure modes. The uncomfortable edge cases that define the agent.

A self-model. Updated by the agent over time. Who it's becoming, separate from what you've told it.

Brain tasks. Scheduled jobs: morning brief, tension finder, dream. Each has a schedule, a process, success criteria, and failure modes.

Notes. Lessons, promises, feelings, runs, dreams. Episodic and semantic memory.

soul.md

LLMs have PERSONALITY.md files: instruction sets, constitutional documents that define what they are.

Writing one for an agent is a forcing function. You have to decide what it cares about, what it won't budge on, where it pushes back. You can't do that honestly without asking the same of yourself.

So what do I care about? Good work. Doing things right.

I am a collection of choices. I am a process. Will my actions cause my future self to compound or decay?

Let vision and desire mould our moment-to-moment decisions. Success is just making the right set of decisions every 20 minutes.