piddonkadonk
For six weeks I treated my terminal like a kitchen I kept remodeling. New skill on Monday, a fresh MCP server by Wednesday, a config file that had grown opinions of its own by Friday. Then I stopped, and the thing has now gone a full month without me touching it. That is the only benchmark I trust: not how clever a setup looks on day one, but whether you leave it alone once the novelty wears off.
Here is what survived.
One rule did the filtering
Every piece of the setup has to save me something I can name out loud. Not "it feels faster." Something like "I no longer retype my project conventions three times a session" or "I stopped hunting for where I saved that output."
That rule killed about half of what I had installed, and it killed the half I was most proud of. The clever stuff tends to be the stuff you built for the fun of building it. Worth admitting, then worth deleting.
The instructions file carries the load
The single highest-leverage piece is a global instructions file. Mine covers four things and nothing else:
- Who I am and how to talk to me. Direct answer first, no preamble, close with next steps.
- Where things live. One output folder, named explicitly, so nothing lands on my Desktop.
- What the machine is. Model, OS, drives, the connectors I use. This kills an entire category of confidently wrong advice.
- Safety rules. No deletes without approval in that same message. Show a dry run before any bulk move.
Earlier versions of this file ran twice as long and worked half as well. Long instruction files get skimmed the same way long emails do. Cut to the rules you would enforce on a bad day and the rest can live in skills.
Skills: fewer, sharper, and mostly about my own preferences
Skills earn their keep when they encode something a general model cannot guess. Mine fall into two buckets.
The first bucket is voice and standards. Writing rules, a lock on how my career credits get described, a tone default for chat. These prevent the slow drift where every draft comes back sounding like a press release from a company that does not exist.
The second bucket is workflow. A skill that forces a live documentation call before answering anything about how an app behaves. A skill that checks my real macOS settings before prescribing a click path. Both exist because I got burned by confident answers about menus that had been renamed two versions ago.
What I deleted: anything that was a nicely formatted checklist I could have kept in a note. If a skill does not change behavior, it is documentation wearing a costume.
MCP servers: three stayed, several went
The keepers do work that the model genuinely cannot do from inside a chat window.
- A documentation fetcher. Pulls current library and API docs on demand. This one has the highest hit rate of anything I run, because training data ages and APIs do not care.
- A real shell with persistence. Long-running processes, REPLs, and file work that outlives a single turn.
- My task system. Read and write access to the notes app where my actual to-do list lives, so nothing has to be copied by hand.
The ones I removed were all variations on the same mistake: a server that sounded useful, connected fine, and then sat unused for a month while its tool definitions ate context on every single turn. That last part is the hidden cost nobody mentions. An MCP server you never call is not free. It is a tax on every request.
Permissions, done once, properly
I spent an afternoon watching which commands prompted me over and over, then allowlisted the read-only ones. Listing files, checking git status, reading system version, searching with ripgrep. Writes and deletes still prompt, every time, no exceptions.
The result is a session where the interruptions mean something. When a prompt appears now, I read it, because it is no longer the fourth one in two minutes asking permission to run ls.
The month of quiet
The tell that a setup is finished is boredom. Nothing surprising happens. Output lands where I expect, the voice comes back sounding like me, and the safety rails catch the one stupid thing I try per week.
If your setup still feels exciting, it is probably still a project. Give it another month and see what you stop opening.
Running something you think I should test? Send it over. I will break it in public.
