A field guide for your AI's mind.
Evolve is a small folder of reading that teaches your AI assistant to check its work, say how sure it is, and learn from its mistakes, instead of confidently guessing.
Your AI is brilliant. It is also, sometimes, a confident guesser.
If you use an AI assistant (Claude, or one like it), you have probably seen it: an answer delivered with total confidence that turns out to be wrong. Not because the AI is careless, exactly. Because nobody ever taught it the habits that careful people live by: check before you claim, say how sure you are, write down what you learned.
Evolve is that teaching. It is not an app and not a program; it is a "skill": a small folder of plain text files (more than 21,000 words of distilled working discipline, counting the core, its reference manuals, and the agent team), which sounds like homework until you remember who it's for: an evening's read for a human, a couple of seconds for an AI. Think of it as the well-worn handbook you would hand a bright new colleague on their first day. Your AI reads it, and acts from it.
You can open every file yourself and read exactly what it says. There is no code in it, nothing that runs, nothing hidden. It is reading material, for a very fast reader.
The same question, two temperaments.
The difference is not intelligence; it is discipline about the truth. (The dialogues are illustrations, not transcripts.)
Four specimens from the collection.
Inside the skill these are called "capsules": twenty short, named rules with examples and limits. A few of them, translated to plain language:
Says how sure it is
"I checked this," "I'm inferring this," and "I'm guessing" become three different sentences, so you know which answers to lean on.
Proof before "done"
It never claims something works without having actually seen it work. "Should be fine now" gets replaced by "I tested it: here's the result."
Three strikes, then rethink
When something fails three times, it stops grinding the same move, steps back, and rethinks (or asks you) instead of burning your time.
Writes its lessons down
Mistakes that cost you time get recorded with what triggered them and how to avoid them, so the same one isn't repeated next week.
- 01The Loop: six steps every serious task runs through: recall, frame, investigate, act, verify, write down.
- 02Twenty-one capsules: named habits like the four above, each with an example and its limits.
- 03Hard constraints: lines that never move: never invent facts, never claim unverified success, never bend a test to pass.
- 04The Laws of AI, and two full manuals: eight foundation laws for honest inheritance, applied by one manual on reasoning and one on writing code.
- 05Memory guides: how to keep notes and lessons across conversations, where the AI's setup allows it.
Designed to travel light.
An AI assistant has a limited attention span: a working memory that must hold your conversation, your documents, and everything it has been told. Every page you hand it takes room away from your actual work. A guide that hogged that space would make your AI worse, however wise its advice.
So Evolve is organized the way a naturalist packs: a slim core that the AI reads whole and carries everywhere, and the thick reference volumes left in the pack until the terrain calls for one.
The core is the only part your AI carries at all times (a couple of seconds' reading for an AI, a small fraction of its attention span), and it tells the AI exactly which volume to pull down, and when. The references, about four-fifths of the total text, stay on the shelf until then: the coding manual comes out only when your AI writes code, the memory guide only when it files its notes. Nothing rides along uninvited, and your conversations keep their room.
Three steps. The third one is a sentence.
You don't install this the way you install software; your AI installs it for you. You just hand it over.
Download the zip
One small file, saved wherever you like. That's the whole skill.
Drop it into a chat
Attach the zip to a conversation with your AI, like you'd attach a photo to a message.
Ask, in plain words
Type one sentence and the AI reads the instructions inside and does the rest.
"Please install this skill."Evolve is free and stays free. A donation unlocks nothing, because everything is already yours; it just says thanks.
Read this before you expect magic.
A page like this usually promises the world. This one won't, because the skill itself forbids it: its first hard constraint is never claim more than the evidence supports. So, plainly:
What Evolve is: working discipline distilled from real, daily use: habits observed to change how an AI talks about certainty, verifies its claims, and recovers from errors. This release was reviewed adversarially before publication, and the skill keeps an open amendment record of its changes, like a proper reference book.
What it is not: a magic upgrade for every situation. The first controlled before-and-after test (July 2026, published in full in the repository) found that on single-question traps, modern Claude models already behave well without it: the measurable effect there was roughly zero. The skill's claimed value lives in long working sessions, where drift and shortcuts creep in over hours; that part is not yet measured, and the skill says exactly that about itself, in writing, on its own pages. A guide that demands honesty from your AI owes you the same.
The best test is the practical one: install it, then ask your AI something you know it might get wrong. Watch how differently it answers.
Accuracy is not free. Neither is being wrong.
Catching a false claim costs something. The AI has to slow down, check itself, and in the full agent version actually search for sources and review its own work before it answers. That is real effort, and it uses more of the AI's capacity than a quick confident guess. Evolve does not hide that cost; it hands you the lever and lets you decide where to stand.
Here is what the checking buys, from our first published test (ten topics, every transcript kept in the repository, and re-runnable yourself):
In ordinary chat (the reasoning discipline, no agents): the model made noticeably fewer verifiably-false claims with the skill than without, roughly half as many in this test.
In the full agent version (where it can fact-check and re-review its own work): across all ten topics, the answers it finally shipped carried zero claims we could prove false, where the naked model's carried several.
The honest counterweight, stated plainly: this is a small test, a demonstration and not a guarantee. And the checking is not free. Every claim the AI verifies for you costs more of its effort than a bare guess would, and the full agent version, which goes and reads the web, costs more still. Think of it as the price of not doing that fact-checking by hand yourself: the work does not vanish, it moves from your desk to the machine.
So the choice is yours, and it is a real one: save the effort and let the occasional false claim through, or spend the extra tokens to stop them. A casual question and a contract you are about to sign sit at opposite ends of that lever. Evolve does not decide for you; it makes the trade visible and lets you set it.
Nothing up our sleeves; it's all on GitHub.
The entire skill is developed in the open, in a public repository: every file the zip contains, the install tooling, and the amendment records that document each change and why it was made. If you want to read every word before letting your AI near it, please do. That's the point.
github.com/Diftic/evolve-siteThe skill, its history, this page, and the zip: all in the open.If you use Claude Code, the repository doubles as a plugin marketplace. One copy-paste line installs the skill permanently: it registers this repository as a plugin source, then installs from it:
claude plugin marketplace add Diftic/evolve-site && claude plugin install evolve@evolveNot technical? Skip this chapter without guilt: the zip above is the same skill, and your AI handles the rest.
The person behind it.
Lars V., online Mallachi, is a self-taught developer: a gamer whose newly discovered passion for building software became the projects on this page.
His day-job world is industrial inspection and compliance for Onix.com, and his background comes from over a decade in oil & gas and a bachelor's degree from NTNU in Technology Design & Business Management. It is a world where a report is not finished when it is written but when it survives an audit against the standard. Evolve applies that habit to AI: if an inspector's claim needs evidence behind it, an AI's claim should need no less.
The rest of his toolbox says the same thing from different angles. He builds automation pipelines, packages his AI tooling so anyone can install it in minutes, and publishes his experiments with the results attached, including the unflattering ones: the near-zero test result in chapter six is on this page because he put it there.
Off the clock the gamer and the developer meet in the middle: he contributes to the Wingman-AI project, writing AI integrations for Star Citizen and Satisfactory, systems that only work if you deeply understand each game's logistics pipelines and the data that drives them.
What he offers, in one sentence: the discipline of a compliance field applied to the newest tools there are, built in the open and documented so you can check the work. If that is the kind of person you want on a problem, the profiles below are where to find him.
github.com/DifticProjects, experiments, and everything Evolve: find him here. linkedin.com/in/lars-erik-vågenThe professional record, in full.