Written by beginnersblog1 on X
What Exactly Is Autoharness?
Suppose you hire a new employee. Every time you correct them, they write it down in a small notebook. Next time the same situation comes up, they check the notebook first, instead of asking you again.
That is basically what autoharness does for your AI assistant.
It watches your coding sessions quietly in the background. When it notices you correcting something, or explaining something, it writes a small note for itself. Next time, it already remembers.
You don't set it up. You don't write any rules. It just learns from the normal work you are already doing.
How Does It Actually Work?
I will break this down into 5 simple steps. It's easier to see it than to read it, so here is the whole thing laid out first:
Now let me explain each step in simple words.
Step 1: It watches. While you work, it quietly keeps an eye on the conversation. What you asked, what mistake happened, what correction you made.
Step 2: It thinks every few turns. After every 3 turns of conversation, it stops and checks something. Do I already have a note about this? Or is this something new?
Step 3: It decides what to do. Maybe it writes a brand new note. Maybe it updates an old one. Maybe two old notes are saying almost the same thing, so it combines them into one. This is the part I like the most. It's not just piling up notes, it's actually cleaning up as it goes.
Step 4: A stricter check has to be approved. This is the part that made me trust it. The AI does not get to save anything by itself. A separate, much stricter checker looks at it first. Is this safe? Is it written correctly? Does it explain why it's making this change? Only after passing all this, it gets saved.
Step 5: If it's not used, it gets removed. Nothing stays forever just because it exists. If a note is never actually used again, it eventually gets moved into an archive folder. Not deleted, just moved aside. You can always bring it back.
Why Does This Actually Matter?
Here is the honest reason I care about this.
You already have the patience to correct your AI assistant once. What you don't have is the patience to correct it 10 times, for the same mistake, spread across 2 months.
Autoharness is built on one simple idea. The correction you make today should not disappear tomorrow. It should be remembered.
There is a number I found in the project's own page that stayed with me. On a coding test called CORE-Bench, the exact same AI model went from a 42% success rate to 78%, just by changing the setup around it. Not the model. Just the setup. That's a huge jump, for the same brain.
Most of us never build that better setup ourselves, because it takes time. Autoharness is trying to build it for you, quietly, in the background.
Let Me Show You A Simple Real Example
Say you are working on a project that uses pytest for testing. One day, Claude Code runs unittest instead, by mistake. You correct it, like you normally would, and go on with your work.
That one correction now goes down one of two simple paths. Here is what that looks like:
A few turns later, the background process wakes up and looks at what happened. It checks, do I already have a note about testing in this project? If yes, it updates that note. If no, it writes a fresh one.
Once it's approved, you will find a small folder sitting right inside your project:
.claude/skills/testing-with-pytest/
SKILL.md
.ledger.jsonl
.sidecar.json
references/evidence-21cd22cc.md
Open the ledger file, and you will see something like this written in plain English:
"User asked about the correct command to run tests in this repo" then skill created. "User pointed out that unittest was being used by mistake, corrected to pytest" then skill updated.
Look at that second line closely. It did not create a brand new file next to the old one. It corrected the same file. That's the whole point. Next time testing comes up in that project, it already knows. You don't repeat yourself.
How To Actually Install It
Install
Requires python3 on your PATH - autoharness runs entirely as Python (zero third-party dependencies); its hooks and MCP server won't fire without it.
/plugin marketplace add tigerless-labs/autoharness/plugin install autoharness@autoharnessRun /reload-plugins (or restart Claude Code), then check it's live:
/plugin # autoharness@autoharness - enabled/mcp # stage_skill — connectedZero config. It now watches your sessions and lands learned skills into .claude/skills/ in the background. Cadence and lifecycle thresholds are tunable - see Configuration.
Update
/plugin marketplace update autoharness # pull the latest release/plugin update autoharness # update the installed copy/reload-plugins # apply in the running session/reload-plugins re-arms the hooks and the MCP server in place; restarting Claude Code works too. The installed copy is cached by the version in plugin.json; releases always bump it, which is what makes the update take effect.
Uninstall
claude plugin uninstall autoharness@autoharness # stops the hooks + MCP serverclaude plugin marketplace remove autoharness # optional — also drops the install sourceWhat I Liked, And What You Should Know Before Using It
What works well:
No manual setup needed, since it learns from work you're already doing. It never touches anything you wrote by hand. Every decision it makes is written down in plain text, so nothing feels hidden. No background service running all the time, and no test score needed to decide what's useful.
What to keep in mind:
It only works inside Claude Code, nowhere else. It learns from real corrections. So a quiet week with very little conversation means it doesn't learn much that week. There's no fancy interface. Everything is plain text files. Good for transparency, but you do have to open a file if you're curious what it learned.
A Few Settings, If You Ever Want To Change The Pace
You don't have to touch any of this. Default settings work fine. But here's what you can adjust if you want.
My Honest Take
I don't think this tool is going to change your whole workflow overnight. It's a small tool, solving one small annoyance.
But small annoyances, repeated 30, 40, 50 times a month, are exactly the kind of thing worth fixing quietly in the background.
If you use Claude Code regularly, and you catch yourself explaining the same project rule again and again, this is worth the 2 minutes it takes to set up.
And if you do try it, I would genuinely love to know what the very first note it writes for you turns out to be. That first note usually tells you something honest, about what you've been quietly repeating yourself on, without even noticing it.
