Most people using AI tools are stuck in a manual loop: prompt, test, tweak, repeat. It's slow, it's tedious, and it's the wrong way to scale.

There's a better approach — one where your AI skills literally improve themselves overnight, without you lifting a finger.

It comes from an idea originally developed for training machine learning models. The concept is simple: instead of a human testing and refining a process, you let a team of AI agents do it autonomously. One agent runs the task. Another evaluates the output. A third mutates the prompt if the result wasn't good enough. Then it repeats — hundreds of times — until performance peaks.

This is called autoresearch. And it works just as well for your Claude Skills as it does for training LLMs.

What Autoresearch Actually Is

Think of it this way: you have a skill — say, a proposal generator, a LinkedIn post writer, or a market research tool. Right now, if the output isn't good, you go in and manually edit the skill's instructions. That takes time and judgment, and you only do it when something goes obviously wrong.

Autoresearch flips this. You define what "good output" looks like — your evaluation criteria — and then let agents run the skill, grade it, and rewrite the instructions automatically. It loops until the skill meets the bar you set.

The result? You wake up to a skill that's measurably better than it was the night before.

The Three-Part System

Here's how to implement autoresearch for any Claude Skill:

1. The Generator Agent

This agent runs your existing skill. It produces outputs — whether that's proposals, emails, analyses, or diagrams. It doesn't know or care whether the output is good. It just generates.

2. The Evaluator Agent

This is where the intelligence lives. You give this agent your quality criteria: What makes a great output? What are the failure modes? It scores each output — pass or fail, or on a scale — and explains why.

3. The Mutator Agent

If the output fails evaluation, the Mutator rewrites the skill's instructions (your SKILL.md). It identifies what went wrong and makes targeted changes. Then the Generator runs again. The loop continues.

The key insight: your program.md (the agent's instruction file) is the equivalent of the ML training script. Your SKILL.md is the model being trained. The evaluator is your loss function.

What This Looks Like in Practice

Let's say you have a skill for generating client-facing project proposals. Your evaluator criteria might be:

You run autoresearch overnight. By morning, the skill has gone through 40+ iterations. The instructions have been tightened, edge cases have been handled, and the output consistently passes all three criteria.

You didn't touch it once.

The same logic applies to any skill: content generation, lead research, email writing, competitive analysis. Wherever you have a repeatable process with a definable quality standard, autoresearch can improve it autonomously.

Progressive Updates: The Simpler Version

If full autoresearch feels complex to set up, there's a lighter-weight version worth implementing immediately: progressive updates.

Build a rule into your skill that says: "Whenever the user tells you not to do something, add that as a new rule in these instructions."

Now every time you correct the skill mid-conversation, it learns permanently. It won't make the same mistake twice. The skill self-corrects based on real usage, not hypothetical test cases.

Over weeks, a skill with progressive updates becomes dramatically more reliable than one you wrote once and forgot about.

Why This Changes Everything

Manual prompt engineering is a bottleneck. The better you get at it, the faster you can improve your workflows — but you're still the limiting factor.

Autoresearch removes you from the loop. It lets your AI infrastructure improve at machine speed, not human speed. The skills you're using in six months can be far more capable than the ones you wrote today, without you having to do the work of improving them.

The compounding effect is significant. Every skill that gets better makes every workflow that uses it better. And every workflow that improves frees up more of your time to build new ones.

That's the real leverage of AI — not just doing tasks faster, but building systems that get better on their own.

Getting Started

You don't need to implement full autoresearch on day one. Start here:

1. Pick your most-used skill. What process do you run through Claude most often?

2. Define what "good" looks like. Write 3–5 specific criteria for a successful output.

3. Add a progressive updates rule to the skill immediately — this costs you nothing and starts working right away.

4. When you're ready, build the evaluator agent around your criteria and automate the loop.

The goal isn't perfect AI. It's AI that keeps getting less imperfect — automatically.