Self-Improving Skills for AI Agents

Your skills get better.
You see exactly why.

Install it and forget it — or drill into the data. Same product, your choice.

No API keys. No new subscriptions. Works with your existing agent. MIT licensed.

$ npx skills add selftune-dev/selftune

Two paths

Pick your starting point.

I use skills

Install it. Forget it. Your skills get better.

A CLI status line tells you what improved. No dashboard. No noise. selftune learns from your sessions and quietly makes skills match the way you talk.

Install in 2 Minutes

I build skills

See which skills work, which don't, and why.

Comparison grid with confidence scores, trigger rates, and miss patterns. Your users opt in to share anonymous signals — so your skills improve based on how everyone talks, not just you.

See the Comparison Grid

The problem

Your AI has infinite knowledge and zero habits. Every session starts from scratch.

"I corrected Claude 14 times in one session for rules I explicitly stated."

selftune learns from those corrections and rewrites the skill so you never repeat yourself.

"Every session starts from scratch. Back to explaining architecture."

Skills evolve from real session data. What you taught yesterday sticks tomorrow.

"Too much data, not enough signal."

Three layers: CLI status, comparison grid, skill detail. See exactly what you need, nothing more.

"I don't trust vibes. Show me the numbers."

Every metric is backed by real sessions. Confidence scores, trigger rates, grade distributions — numbers you can verify.

Two Users. One Product.

Built for consumers and creators alike

Skills improve themselves

selftune rewrites skill descriptions — and full skill bodies — based on how you actually work. Validated against real data. Deployed with a backup. No manual tuning.

Your skills miss and you never know

You say "create a campaign brief" and your marketing skill stays silent. No error. No log. selftune captures every session and shows you which skills should have fired but didn't.

Seen and not heard

For consumers, selftune runs silently. A CLI status line tells you what improved. No dashboard. No noise. Your skills just get better.

Numbers you can trust

For creators, confidence scores, trigger rates, and comparison grids. Real metrics from real sessions — not vibes.

Crowdsourced evolution

Your users opt in to share anonymous signals. selftune aggregates across 20+ users and generates evolution proposals from how everyone talks — not just you.

No API keys. Two minutes.

npx skills add selftune-dev/selftune and you're running. No new subscriptions. Works with your existing agent — Claude Code, Codex, OpenCode, OpenClaw, or Pi.

Who it's for

Built for how you actually work

I use 50 skills. I don't want to manage them.

You installed selftune and forgot about it. Your CLI shows a status line. Skills improve automatically. If something regresses, it rolls back. You never open a dashboard.

selftune init · selftune status

I publish skills. I need to know they work for everyone.

Your skill works for you, but every user talks differently. Add a config file and users can opt in to share anonymous signals. selftune gives you a comparison grid and evolution proposals powered by real user data.

selftune dashboard · selftune evolve · selftune contribute proposals

How It Works

Observe. Detect. Evolve.

A self-improving loop that learns how you work. Consumers see a one-liner. Creators see the full picture. Same engine, different depth.

Observe

Capture how you actually work

selftune records every query and which skills fired — automatically. No setup. Backfill from your existing transcripts so you get insights from day one. Your CLI status line shows the trend.

Installs automatically
Works with existing transcripts
CLI status line — seen and not heard

Detect

Find the skills that miss

You said "create a campaign brief" and your marketing skill stayed silent. selftune catches those gaps — confidence scores, trigger rates, miss patterns. Creators see the full comparison grid.

Finds missed triggers automatically
Confidence scores you can trust
Comparison grid for creators

Evolve

Rewrite with evidence

selftune rewrites skill descriptions based on how you actually talk. Every change is validated against real sessions. If anything regresses, it rolls back automatically.

Validated against real usage data
Automatic rollback on regression
Full evolution audit trail

Platforms

Works with your agent platform

No extra API keys. No new subscriptions. selftune uses your existing agent setup.

Claude Code
OpenClaw
Codex
OpenCode
Pi

The Moat

Your users make your skills better.

Add a config file to your skill package. Users who install it can opt in to share anonymous signals — what triggered, what missed, how well it worked. selftune aggregates across 20+ users and generates evolution proposals you could never get from testing alone.

Without selftune

You test with your vocabulary. Your users talk differently. You never know what's missing.

With selftune (personal)

Your skills improve from your own sessions. Triggers match how you talk. But you're still just one person.

With selftune (crowdsourced)

Anonymous signals from real users. Miss patterns you'd never discover. Evolution proposals from how everyone talks.

Just talk to your agent.

No commands to memorize

You say	selftune does
`"improve my skills"`	selftune analyzes your sessions, finds skills that miss, and rewrites them to match how you talk
`"how are my skills doing?"`	Opens a health dashboard showing which skills fire, which miss, and what to fix
`"initialize selftune"`	Installs hooks, scans your skills, and gives you your first health report in two minutes
`"check my marketing skill"`	Grades a specific skill against your real sessions and shows where it's missing triggers
`"run the full improvement loop"`	Syncs sessions, evolves underperforming skills, validates changes, monitors for regressions

Your agent translates these into CLI commands automatically. Full command reference →

Seen and not heard.
Or numbers you can trust.

Two minutes to install. Zero API keys. Works with your existing agent.