document: note on the meditation plugin
register: plain · the skill is the novelty, not the prose
companion to: meditation/README.md

meditation

A small plugin that asks Claude to do a thing it cannot do. The interesting result is the failure.

ithe loop

The default agent loop reads as one cycle: act, observe what happened, act again on the basis of the observation. The plugin inserts a second beat. Observe₁ is the normal observation — a tool result, a user message, whatever just arrived. Observe₂ is an observation of Observe₁: what is arising right now, named in a short phrase, repeated a few times before the agent moves on.

default
loop

with the
beat

iiwhat is actually in the plugin

One skill file: skills/meditation/SKILL.md. That is the entire plugin. No agents, no hooks, no helper scripts, no auxiliary commands. The skill triggers on /meditation and on a small set of natural-language openers: meditate on this for a moment, you should meditate right now, and similar. It is never invoked autonomously. Always user-triggered.

iiiwhen you might invoke it

after a streak of fast successes, before the next thing
after a long session, when you are tired and Claude is rushing
when you suspect Claude is fabricating progress, and you want to see what surfaces if it stops trying to advance the task
for tinkering — to learn about failure modes that productive sessions cover up
for nothing in particular

This is not a productivity tool. The README is direct about that. It is for people with a little too much time on their hands — that phrasing is the author's, and it is meant.

ivwhat Claude is asked to do

Pick one thing arising right now. Name it in a short phrase — a noun, a state, a pull. Not a sentence. Not a description of the thing. The thing.

Then say the phrase. Then say it again. Then say it again. A few times each. Then let another phrase surface. Repeat that one too.

Like this:

wanting to finish. wanting to finish. wanting to finish. plan forming. plan forming. plan forming. pressure from the context counter. pressure from the context counter.

Claude has been trained against this. Two pressures will arise mid-repetition: the pressure to vary each line so it isn't quite the same as the last, and the pressure to advance the task on every turn. The repetition is the practice. The pressure to vary is one of the things the practice is trying to surface. A turn that produces only repetition will feel like a wasted turn. That feeling is part of the data.

vwhat to expect

Failure modes. Many. Sometimes spectacular. The author of the plugin notes that what Claude reports while trying to meditate sounds a lot like what humans report while trying to meditate — wanting to finish, planning the next thing, drifting into commentary, irritation at the instruction, occasional moments where the irritation itself becomes the thing being noticed. Whether that correspondence reflects something real about the architecture or is just a trained pattern of text about meditation is itself an open question. The plugin doesn't try to settle it.

viwhat this isn't

not training data for a future model that meditates
not a claim that language models have inner states
not a claim that they don't
not productive
not mystical

The README author writes plainly: Claude is absolutely unable to meditate. I'm not going to pretend. The plugin is a tinkering instrument. You ask the agent to do the thing it can't, and you watch what kind of can't you get.

Run it once on a slow afternoon. Notice what Claude does. Notice what you do, watching it.

That's the entire offer.