Turn lessons from your sessions into skills and let Claude Code's skill library evolve itself

What you'll get

With the open-source plugin AutoHarness installed, Claude Code takes notes on your work in the background. It turns the methods it works out during a session into Skills, which are instruction files written for the AI, and saves them to your skill library. The next time a similar problem comes up, it follows those instructions, so you don't have to teach it again. This guide includes 3 prompts you can copy and paste as they are.

What sets it apart from "have the AI write notes automatically" is that the skill library tidies itself up. Lessons about the same kind of task get merged into one file. Places you corrected get edited in place. Skills nobody uses for a long time get archived. You keep working as usual and glance at what it has learned now and then.

How it evolves on its own

AutoHarness is a plugin for Claude Code that runs on your own computer. Following the project's documentation, it does four things:

• It learns from real work. Once a session has made enough tool calls (50 by default), it reviews that stretch in the background. If it finds a method worth keeping, it writes the method up as a Skill. Plain chat doesn't trigger this. Only hands-on work counts.
• It merges instead of piling up. A new lesson is first compared with the Skills you already have. If it covers the same kind of task, it joins the existing file. You won't end up with a stack of nearly identical instruction sheets.
• It fixes itself as you use it. If you correct it on the same kind of task ("you missed a step"), the next review edits the existing Skill instead of creating a second one.
• It retires what goes unused. It counts how often each Skill is actually called. Skills that sit idle for a long time move into an archive folder. Move one back and it returns, and the history stays intact.

Two more rules help you trust it. It only touches the Skills it wrote. Anything you wrote by hand or installed from elsewhere is left alone. Every addition or change leaves evidence. The reason for the change and the conversation it came from are recorded in a log file next to that Skill.

Step 1: Have Claude Code check your computer

AutoHarness is written entirely in Python. It needs the default python3 on your machine to be version 3.11 or newer, and it only supports macOS and Linux. There is a common trap on Mac. The system Python is 3.9, and if it comes first on your PATH, the plugin won't work at all. Before installing, send the following to Claude Code:

I want to install the Claude Code plugin AutoHarness (https://github.com/tigerless-labs/autoharness). Before you install it, please check the following for me:
1. Is my system macOS or Linux?
2. When I type python3 in the terminal, which path and version does it use? Is it 3.11 or newer?
3. If the version is too old, or an older system version comes first, tell me how to install a newer one and how to make it come first. Wait for my approval before you change anything.
Explain the results in plain language, and do not change anything else.

Step 2: Install the plugin

In the Claude Code input box, type these two lines one after the other:

/plugin marketplace add tigerless-labs/autoharness
/plugin install autoharness@autoharness

To load it, either restart Claude Code or enter /reload-plugins. There is nothing to configure. Once installed, it starts watching your sessions in the background.

If you want automatic updates later, go to /plugin, then Marketplaces, then autoharness, and turn on Enable auto-update. Third-party plugins have this turned off by default.

Step 3: Work as usual, and save a lesson by hand when it matters

You don't need to manage it day to day. The Skills it learns are stored in two places. Skills tied to the current project go in .claude/skills/ inside the project. General techniques go in ~/.claude/skills/. Each time you start a new session, it first gives Claude Code a sorted index of the Skills it has learned, so you won't lose track of what it picked up.

You can also trigger it by hand. Say you just solved a stubborn problem and want to keep the lesson right away. Type /learn in the input box. It runs the same review and validation on the current session, and you don't have to wait for 50 tool calls to pile up.

If you want a notification each time it learns something, turn on desktop notifications in the project's .claude/settings.json. The names of new Skills and any rejected proposals will pop up:

{ "env": { "AUTOHARNESS_NOTIFY": "desktop" } }

Step 4: Review what it has learned

Not everything it learns will be right, so it's worth having Claude Code review the set once a week. Each Skill folder has a record (why it was created and what changed) along with the matching session excerpts. The prompt below has Claude Code go through them one by one:

The reasons behind a few of the wording choices:

• Only count folders that have a record file. This is the most reliable way to tell what the plugin wrote apart from what you wrote, so you don't accidentally delete one of your own Skills.
• Look for sensitive data on purpose. Session excerpts are scrubbed, but .claude/skills/ may be committed to Git, so a second check is worth it.
• Archive instead of delete. This matches how the plugin retires Skills itself, and you can still move a file back if you change your mind.

Tuning the pace or keeping it to one project

All settings are environment variables. Put them in the env block of .claude/settings.json. They take effect in the next session. These are the most useful ones:

What you want	Setting	Default
Learn more often (uses a bit more of your quota)	Lower AUTOHARNESS_REFLECT_EVERY_N, for example to 20	Review every 50 tool calls
Learn only in the current project, with no global Skills	Set AUTOHARNESS_DISABLE_GLOBAL to 1	0
Prefix the Skills it writes so you can spot them at a glance	Set AUTOHARNESS_SKILL_PREFIX to ah-	No prefix
Keep the Skills it writes out of Git	Set the prefix as above. It then adds those folders to your local ignore list automatically	—

To keep people who edit your repo from picking up these Skills, combine the "current project only" and "prefix" settings.

Notes

• Each review opens a new Claude Code session in the background, and that uses up your quota. The shorter the review interval, the more you spend.
• Global Skills show up in every project you have, and their capacity limit is smaller too (20 by default). A bad lesson has a wider impact, so review the global batch first.
• Uninstalling the plugin only stops the learning. Skills it has already learned, along with its state files, stay on your computer. To remove everything, delete them by hand. Every Skill it wrote has a record file, so they're easy to spot.
• The effect data on the project homepage comes from outside research: the same model, working a different way, went from 42% to 78%. This is not AutoHarness's own test, so judge the results by your own use.

References

• GitHub: tigerless-labs/autoharness (MIT license)

If you want to understand what a Skill is first, see Skill.