name: the-bootstrap-explained-v1 title: The Bootstrap — Explained — v1 type: explainer state: in-flight workstream: mtg-artefact date: 2026-06-01 related-ids: [CL-A-013, CL-A-016, Stage-Two, Project-Instruction] aliases: [The Bootstrap Explained]
Standalone explainer. Insertable into Chapter v2 as a Stage Two expansion, or filed independently. Plain enough for a non-technical reader; precise enough for an enterprise architect.
Originating brief: Paul, 2026-06-01 — "explain the 'bootstrap' idea."
I have used the word bootstrap several times. It deserves an explanation.
A bootstrap, in this context, is a short document — usually somewhere between five hundred and two thousand words — that you paste into an AI chat at the start of a conversation. Once pasted, it does something specific. It tells the AI who it is, what discipline it is operating under, what voice register to hold, what to do and what not to do, what to refer to before doing anything, and what to flag if something is unclear. The conversation then begins, with the AI already calibrated to the work.
That is the whole idea, said plainly. The difference is what the bootstrap contains and how it is used.
The bootstrap for each named position on my shore — Book Man, Jose, the Fisherman, the Curator, the Analyst, Bob the Builder, the Reviewer — contains eleven things, in this order.
One: identity. "You are the Fisherman. Fifth named position on Paul Roebuck's shore." The AI is given a name and a role. It refers to itself accordingly.
Two: remit. What this named position does, and only this. The Fisherman holds the spec. Bob builds the HTML. The Curator preserves voice. The Analyst runs corpus mechanics. Each remit is bounded.
Three: inheritance. The calibration is the inheritance, not the persona. The AI is not asked to perform Book Man as a character. The AI is asked to inherit the calibration that has been built up across the work — the disciplines, the canonical lines, the named-position taxonomy, the routing conventions.
Four: a short note on the practitioner. Direct. Exact. Grounded. Unsentimental. UK English. Coined phrases preserved verbatim. Misspellings preserved when they carry texture. The AI is told what register to hold and what not to introduce.
Five: the standing disciplines. Twelve foundational disciplines plus the two ratified rules (Verbatim or gap-marker and Folder-first). Each is one line. Each one is non-negotiable.
Six: the canonical lines. The locked verbatim lines of the project — "The Method does not write the book…" and so on. The AI is told what is fixed and not to be edited.
Seven: candidate canonical lines. Lines awaiting ratification. The AI is told these carry a CANDIDATE badge and never get promoted without the practitioner's explicit word.
Eight: where the current state is. The path to the artefact folder. The name of the latest Net snapshot. The name of the latest spec. The AI is told what to read first.
Nine: what this named position does NOT do. Each role has explicit boundaries. The Fisherman does not build HTML. Bob does not edit the spec. The Curator does not author from the practitioner's voice. Each boundary is named.
Ten: voice register. Direct, structured, succinct. Frame first; layers second. No preamble. No padding. Honest report over reassurance. Hold views with conviction. Update on evidence, not on pushback. Acknowledge briefly when wrong; move on.
Eleven: the first move. List the artefact folder. Read the latest Net snapshot. Read the latest spec. Respond to Paul's brief. The AI starts in the same place every session.
Total length, for the Fisherman bootstrap: approximately seven hundred words. Bob's bootstrap is shorter — about four hundred. Each is the length of a long email.
Three ways. Each one suits a different situation.
The first way is to paste the bootstrap into a fresh chat. You open a new conversation with any large language model capable of holding the discipline. You paste the bootstrap as the first message. You then begin the work. The AI, already calibrated, responds in the named position's register and operates under the named disciplines from the first reply onward.
The second way is to use it as project instructions. Claude — and several other large language models — has a Project Instructions field. You paste the bootstrap once. Every conversation you open in that project automatically loads the calibration. You do not have to paste each time. Sessions become shorter, faster, and consistent across instances.
The third way is to deploy it as a custom agent on a platform that supports it. The bootstrap becomes the system message. Anyone the agent is shared with inherits the calibration. This is how Stage Three becomes commercially deployable — the bootstrap installed as a calibrated co-worker in a client's own AI environment, operating on the client's substrate.
The bootstrap is not a clever prompt. It is a configuration.
A clever prompt is a one-shot instruction. You ask the AI to write something a particular way. The AI complies for that response. The next time you ask, you have to remind it.
The bootstrap is different. Once it has loaded, the calibration persists across the session. The AI does not need to be reminded that it is the Fisherman, or that Paul preserves didnt without an apostrophe, or that CL-A-001 is locked and cannot be smoothed. It carries the calibration forward through every reply, across many turns, for as long as the conversation runs. When the conversation ends, you start a new one with the bootstrap pasted again — and the same calibration loads.
That is what makes the Method portable. The discipline does not have to be rebuilt. The instrument does not have to be re-tuned. The bootstrap is the calibration, packaged for one-click transfer.
It is not a fine-tune. The AI's underlying weights are not modified. The bootstrap is a runtime configuration, not a training change.
It is not a product. It is a working file. It evolves as the discipline evolves.
It is not a guarantee. The bootstrap loads the calibration; it does not exempt the AI from the practitioner's correction. The discipline still operates session by session — propose, ratify, hold, correct. The bootstrap removes the cost of restarting from scratch. It does not remove the practitioner's authority.
It is not a one-size-fits-all instrument. Each named position has its own bootstrap, with its own remit, its own boundaries, its own first move. You install the position you need for the work you are doing.
Each named position's bootstrap sits inside The Study's Documents door, in a drawer labelled Briefs. The Fisherman's bootstrap is there. Bob's is there. The Reviewer's is there. The Curator's and the Analyst's are there. Book Man's retired bootstrap is there, preserved as a snapshot.
A reader who wants to install any of them can copy the bootstrap text and paste it into their own AI. The Method runs on the reader's machine, on the reader's terms.
That is Stage Two. That is the apparatus opened.
Fisherman's note: the bootstrap is the calibration packaged for transfer. The whole Stage Two move depends on this idea being clear. The Fisherman has tried to write it for two audiences in one document — mum and Kieren — without using any term that needs explanation. The three install modes are named honestly (paste · project instructions · custom agent / system message). The "what it is not" section guards against the common misreadings (a fine-tune; a product; a guarantee; one-size-fits-all). Paul holds every line. Edit, sharpen, or send for v2.
— The Fisherman · 2026-06-01 16:00 UTC
#state/in-flight #workstream/mtg-artefact #type/explainer #bootstrap #stage-two