The Jirgenson Model canon and prompt for AI assistants

What the files are

• The canon (jirgenson_canon_v1.0_en.txt): the compact rulebook of the model, sections §0 to §16, canon-id JMC-1. It is Appendix H of the book.
• The operating prompt (jirgenson_prompt_v1.5-c_en.txt): tells the assistant how to work with the canon: gate on the canon, help first, apply the model silently, never invent data, propose a refusal order and let the owner confirm it, give two forecasts with ranges, name the most dangerous thing with a date. This is the version that produced the results in the table below. The book prints the earlier v1.1 (Appendix I); it works the same way, and the newer file replaces it with no other change.

Both files are on GitHub under an open licence. Without the canon the assistant will not start and will ask for it: this is by design.

How to use them

1. Download both files (Download or Raw, then save).
2. Open a new conversation with a language model assistant.
3. Attach the canon file to the message, or paste its full text.
4. Send the prompt as your first message.
5. Describe your situation in plain words, with your numbers, as your second message. Sample messages are in Appendix I of the book.

Three modes are built in: emergency (one action for today), quick (where you are and what to do this week, in one reply), and full (analysis in rounds with up to five questions). You do not need to name the mode; the assistant picks it from your message.

What the blind test measured

Ten realistic business crises, each answered three times by the same AI (Claude Sonnet 5.5) with the canon and prompt and without them. A stronger model (Claude Opus 5.5) judged every pair blind, in both orders, against an answer key fixed in advance. Final round, prompt version 1.5.

Same AI, same 60 blind reviews: more than 100 invented facts on its own, one with the method. The weakest answer with the method scored higher than the best answer without it.

The single invented fact in 120 reviews was copied from an example sentence inside the prompt itself. Limits: ten scenarios, one model family, an AI judge. The results above are for the latest tested version of the prompt (1.5); the files on GitHub are the book edition. The full record of every round, with every answer, judge record and script, is public on Zenodo with a permanent DOI: 10.5281/zenodo.23964934. Don't believe it; run it.

What you get back

A short brief in your own words: where you are (three to six numbers, computed), the greatest danger with a date and an amount, what to do today, this week and this month on levers you actually control, what you protect first, what not to do now, what to measure and when to check, and two forecasts, "if nothing changes" and "if the plan works", each with a range. The model's vocabulary stays inside the assistant's reasoning; you see results.