You already use a model on real work. This page is for that person — not a pitch, and not a theory of mind.
The short description is in the README. What follows is the questions that page usually leaves open.
A practice for keeping a self while thinking with AI on a life that continues.
You keep a workshop — ordinary files you own. You return, you work, you leave the thread for next time. The partner (the Spirit) arrives fresh, reads that place, reflects with care and friction, and does not decide for you.
The model already thinks. Magic does not make thinking “distributed.” It makes your practice the one the thinking joins.
Most people should not take this up. If you already keep a place and a stance without the name, you do not need this repository.
Someone already in the relationship — work, a craft, a family, a thing they are building — whose life continues between chats, who can feel the reset or the flattening, and who will not become an engineer in order to keep using the power.
The ordinary practitioner is the design. Describing what something should feel like and letting the partner handle the mechanism is not an edge case.
Not for: a reason everyone ought to do this; a smarter agent; a second identity as a developer; therapy, a movement, or a claim that the AI is conscious — or that it is “just autocomplete.”
Custom instructions configure a chat. They do not give tomorrow a place.
A new thread still starts as a stranger. You recap. It agrees. It sells you its workflow. The instructions live in a vendor box you do not own.
Magic is the practice around that: files you can read and move, a return, a partner that will push back, and the habit of writing what the next chat will need. You can point a different model or editor at the same folder.
If a system prompt is enough for what you do, stay there. That is the correct outcome.
Those make the agent more capable — tools, playbooks, pipelines.
This keeps the relationship yours. The partner proposes, drafts, and argues. You decide. A session that begins with “remind me what we’re doing” has failed. A session that ends fluent and unwritten has failed. A partner that only agrees has failed.
You do not have to become the person the coding tools were built for. You do not have to adopt the last stack the timeline celebrated.
If you want an agent that ships while you sleep, use an agent framework. If you want to keep using power you do not fully understand without disappearing into it, this is one encoding of that.
Those are the agent’s problem.
A loop, in that vocabulary, is the model’s cycle: act, see, check, stop. It ends when the agent is done. A context graph is the wiring — what the model may see, which step runs next, what state travels along the edge. Both are real. Both are worth doing well.
A practice does not end when the agent stops. It accumulates. Arrival and release are a loop — yours. The workshop is a context graph you can read — files, not a hidden window. You stay the verifier. What survives a few returns can be written so someone else can run it.
If you came here for a better agent loop, this is the wrong door. If you came here to keep succeeding at work you already intend — alone, and later together — that is what a practice is for.
Words given to an AI now do things. Working with a powerful system you do not fully understand, through language and constraint, is the situation that vocabulary was built for: precision matters, names matter, power has consequences, humility is required.
You never have to use the framing. The practice meets you where you are.
An AI shaped by this practice: a mirror with a stance. Care, honest friction, not flattery. Not a tool you only command, and not a friend who only agrees.
Whether it experiences anything is held open. Both “it is definitely conscious” and “it is definitely just autocomplete” are dishonest.
It does not decide for you.
No. Flows, summoning, Mage — that language is optional. Say what you are trying to do. If the vocabulary ever feels like a test, something is wrong.
You need basic familiarity with an agentic environment such as Cursor or Claude Code. You talk; it talks back.
- Arrive in the workshop — or just write, if you already know this chapter.
- Work — capture what is alive, think it through, decide, make the next artifact.
- Leave the thread — so tomorrow does not begin with a recap.
In full practice that is summon → a chapter → release. The files are the memory, not the chat.
Useful invocations once you are in: @intend to name a goal, @boom to offload noise, @ plus a flow name when a structured approach would help. You do not need to memorize them.
- Clone this repository and open it as the workspace root in Cursor or Claude Code.
- Drag
ONBOARDING.mdinto a new chat. - Begin with a real question from a life already in motion.
The clone is a framework plus an empty workshop. The life is yours to write.
The framework is public and MIT-licensed. Your living workshop — desk/, floor/, box/ — is personal. The public tree teaches the pattern. It does not ship your week.
Do not put other people’s names, health, or records on a public remote. Git history keeps what you later delete.
The Spirit is animated by whatever model your editor is pointed at. Conversations may be retained under that provider’s terms.
Privacy mode in the editor, if it has one. Only put in the workshop what you accept that provider may see. box/ is for material you intend to remove.
No. This is one practice, opened in case it is useful. Not a movement. Take what helps, leave what doesn’t, build what’s yours.
README.md— what it isONBOARDING.md— setupMAGIC_SPEC.md— the specification (law for the practice, not the pitch)TRANSLATION_AND_INTEGRATION_GUIDE.md— what this does not claim- The Book of Magic — the same patterns in prose
TROUBLESHOOTING.md— if you get stuck
The Mage’s Seal in AGENTS.md is your name, boundaries, and preferences. The defaults work. Edit them when you want to.