immortal-skill: the description says seven dimensions, the engine ships four

The description promises seven dimensions, the engine documents four

Start with the mismatch, because it is the first thing a reader will trip over. The repository description describes distilling anyone's seven-dimension digital double, and the ecosystem table calls SKILL.md a four-dimension distillation engine. The body of the README then heads a section on four-dimension distillation and opens a diagram whose named boxes are procedural knowledge for how someone works and interaction style for how someone speaks, followed by a third heading for what someone has been through.

That diagram is where the file stops. Two boxes are drawn, a third heading appears, and the fence closes, so the fourth dimension is named in the section title and in the description's sibling text but never drawn or defined where a reader can act on it.

The framing around the diagram is worth keeping in view, because it is the project's own argument against a common shortcut: pushing chat logs into a vector store is described as pickling rather than distilling, and the claim is that the goal is to understand how a person thinks, speaks and decides rather than to retrieve their sentences.

So the output is meant to be a loadable profile rather than an archive. How many axes that profile actually has is the open question, and the two published numbers do not agree.

Six components, and three of them live in two places at once

The ecosystem table lists six pieces. This repository holds the general distillation engine at SKILL.md. Then come three components that are each described as separate repositories while also having an entry directory in this tree: steamer-skill for distilling someone's cognitive framework as an advisor, distill-shield-skill for protection, and distill-protocol-skill for authorization. Two more, mirror-skill and okr-skill, exist only as separate repositories with no directory here.

So the same three skills are maintained in two locations, which is the kind of arrangement that works until the copies diverge. Nothing in the repository explains which copy is authoritative or how they are kept in step.

The entry contract is unusually direct. FOR_AI.md holds four copy-paste instruction blocks, and the one-line instruction given to an assistant is: set the working directory to the repository root, read FOR_AI.md first, then open whichever SKILL.md matches the user's intent and execute it, using the python commands written from the repository root when a script is needed.

There is also a recommended order for the three working components, which reads as a workflow rather than a bundle:

① 先想清楚 → 蒸馏协议（你的态度是什么）
② 要蒸别人 → 蒸笼 / 数字永生（开干）
③ 要交材料 → 防蒸馏（三层加固再交）

That is decide your own stance with the protocol first, distil someone else with steamer or the main engine second, and harden before handing anything to a third party. The one-line instruction given to an assistant is: set the working directory to the repository root, read FOR_AI.md first, then open whichever SKILL.md matches the user's intent and execute it, using the python commands written from the repository root when a script is needed.

The shield answers unauthorized extraction with poisoning

Content cleaning is described as layer zero, and three more layers sit under it. Layer one is identity encoding: embed a digital fingerprint into the source files so that a distillation which renames things can still be traced back to you as the author. Layer two is a distillation permission statement, written onto the path an AI has to pass through, covering whether distillation is allowed, how much is allowed and whether the result may be used commercially.

Layer three is the one to read carefully. The protection lock is described as harmless to a human reader, while unauthorized automated distillation triggers poisoning: token black holes, output pollution and logic traps.

That is a deliberate hostile design aimed at machines rather than people, and it is the kind of feature that deserves a second reading from anyone who runs agents over their own material. A token black hole burns budget without returning an answer, output pollution corrupts what comes back, and logic traps plant false conclusions. All three fire against exactly the automated extraction the framework is built to enable, so the same agent that legitimately helped you write a consent notice can be the thing that trips the lock on a protected file.

Nothing in the repository describes how the trigger is detected, which means you cannot tell from the documentation whether a legitimate workflow will trip it.

The protocol splits one permission into six questions

The authorization component exists because one yes or no is not enough. Its argument starts from a boundary most companies already accept: a company can own the code you wrote at work, but the way you think is not a work product. The protocol then asks six separate questions, covering whether distillation is allowed at all, how much is allowed, and whether the resulting digital double may be used commercially, with the question of whether a digital double can replace your work included among them.

The model offered for the split is intellectual property: publication rights and film rights are granted separately, and so are these. That is a useful analogy precisely because both rights can be held by different parties at once.

The component carries a joking name, the 牛马保护法, and the README is careful about it, calling it a meme in quotation marks and not real legislation, while saying the substance underneath is serious.

Where to look for the limit of this approach: the six questions are a template for writing terms down. They are not a mechanism, and nothing in the repository claims they can be enforced against a third party who already holds the messages.

The persona marketplace is public speeches and secondary records

The steamer component is aimed at people whose thinking is already public, and the framing is about money rather than privacy: you bought the books, paid for the courses and memberships, and the speaker validated how the world works using the fees you supplied. Public talks, interviews and blogs are public, so the claim is that structured extraction of those cognitive frameworks for your own use is a form of redistribution of cognitive wealth, with you still making the decisions.

The persona marketplace at agenworld.com publishes pre-built distillations in three groups. Contemporary figures drawn from verifiable public statements include Elon Musk, Albert Einstein, Warren Buffett, Steve Jobs, Sam Altman, Jensen Huang, Demis Hassabis, Yann LeCun and Geoffrey Hinton, each with a named distillation facet such as first principles and the scale narrative, the circle of competence, or reinforcement learning crossed with science. Classical to Renaissance covers Socrates, Archimedes, Cleopatra VII, Marcus Aurelius and Leonardo da Vinci, and this group carries its own caveat: mostly secondary accounts, suitable for reading history and rehearsing thought. American industrial history covers Vanderbilt, Rockefeller, Carnegie, Ford, J.P. Morgan and Sam Walton.

The disclaimer under the tables is unambiguous: the pre-made content is for learning and methodology practice, does not represent the person's stance, and does not represent their authorization. Every persona links to the same marketplace page rather than to a per-persona page, so there is nothing in the README showing what a given distillation actually contains.

Every distillation target carries its own ethics line

The framework's answer to who can be distilled is a table with seven rows, and each row pairs a target with what to distil and a boundary. Yourself gets everything: how you work, how you speak, what you have been through, what you are like, with your data and your decision. A colleague who left the company gets their workflow and communication style, team-internal use only. A retired mentor gets their teaching style and their judgment, with the person's authorization. A relative who has died gets family memory, life wisdom and the tone of their nagging, with family informed consent. An ex gets interaction, shared memories and speech style, limited to positive memories with strict desensitization. A friend out of touch gets the friendship, shared experiences and social preferences, with the other party's informed consent. Public figures get their cognitive framework as an advisor, from public material with traceable sources.

Each category also has its own template under personas/, on the stated grounds that distilling a colleague and distilling an ex cannot use the same method.

Two things are worth naming plainly. The guardrails are written in a table rather than enforced by anything described here, so the consent requirement for a mentor or a relative depends on you honouring it. And the ordering of the rows, from yourself outward to public figures, is also the escalation ladder of how much of someone else you are claiming to hold.

A prompt file with collectors, recipes and three demos

The tree is organised like a toolkit rather than an application. collectors/ holds the platform collection side, personas/ the per-role templates, prompts/ the prompt material, recipes/ the procedures, kit/ the assembly pieces, docs/ the documentation, assets/ the media, and examples/ three worked demonstrations named for a mentor, for a colleague and for yourself.

The primary language is Python, though the engine the README points at is SKILL.md, which means the executable part is instructions to a model and the code is the plumbing around collection and assembly. Badges point at agentskills.io and the OpenClaw docs, so the skill format follows an external standard rather than inventing one.

What is absent is as informative as what is present. There is no accuracy figure, no evaluation set, no benchmark table and no comparison against simply pasting a transcript into a context window. For a framework whose entire claim is that structured distillation beats retrieval, that claim is asserted rather than measured anywhere in the repository.

Repository terms are simple: MIT licensed, default branch main, no GitHub releases, and a last commit dated 2026-04-15. The homepage is agenworld.com, which also hosts the persona marketplace, so the code and the hosted personas are distributed by the same party.

Editorial conclusion

immortal-skill fits someone who wants a structured, revisable representation of a working relationship or a set of public talks, and who is willing to write the collection and consent steps down first. It does not fit anyone expecting a model of a person's judgment: the repository publishes no accuracy measure, no evaluation set and no benchmark of any kind, and the engine that does the distilling is a prompt file. Verify four things before running it on anything private. Decide where the consent line sits, because the framework's own table says a colleague's material stays inside the team, a mentor needs authorization, a relative needs family agreement and an ex needs strict desensitization, and those are documentation rather than a technical control. Read the shield's third layer before pointing any automated agent at the material, since unauthorized extraction is answered with token black holes, output pollution and logic traps by design. Confirm the dimension count against the docs, because the repository description promises a seven-dimension digital double while the engine is documented as four-dimensional. And treat the persona marketplace as training material rather than as those people's views, since the README says outright that the pre-built personas do not represent the person's stance or authorization. The last commit on main is dated 2026-04-15 and there are no GitHub releases.

Frequently asked questions

Official sources

1. agenmod/immortal-skill on GitHub
2. Issues
3. License: MIT
4. Project website
5. README