Give Your AI Models Proven Skills

Skills are not just SKILL.md ✗ files. You can uplift any LLM model's performance by up to 20% with lab-tested ✓ skills.

or start with existing skillUpload Skill
The problem

You create SKILL.md for AI models.
But never test them.

Don't trust skills blindly. They can degrade LLM models and you won't know it.

SKILL.md
1---
2name: code-reviewer
3description: Reviews a diff and flags
4  real defects.
5---
6 
7Read the whole diff before judging any part of it.
8Flag correctness bugs first, style never.
9Keep every note under two lines.
10Do not approve a change you cannot explain.
SKILL.md savedtests run: 0 · score: · shipped ✓
Does this skill improve
the model's performance?
never tested
Is this version
better than the last?
no skill versions exist
Did you run evals
on your skill?
never ran them
Does it have
self-improvement capability?
nothing like that
Will this work well
in all harnesses?
no way to know
Can you prove this is
the best skill version?
never ran benchmarks

You shipped it anyway.

The lab

Create, test and run benchmarks on your skills

Skills can uplift an LLM model by 15%, but only tested skills do. The lab tests every draft, fixes what fails, and saves it as a new version — v1, v2, v3 — until the skill is good enough to ship.

01 · Measure

Every skill version gets a score.

still improvingready to ship

The lab creates the eval cases, then runs tests on your target model and scores your skill's performance. It keeps refining and testing the skill until the score passes 90. Only that version ships.

02 · The loop

Test. Refine. Re-run.

The lab creates a self-improvement loop: it runs the skill, tests it against the eval cases, and fixes what fails. Each pass becomes a new version. The loop continues until the lab hits its goal.

  • +40points better than the model with no skill
  • 10/10eval cases passing before it ships
  • v10the first version to pass 90. Earlier versions never leave the lab
03 · Benchmark

Your tuned skill vs the base model

Base modelTuned skill

No skill ships on trust. Every release is benchmarked on custom eval cases — base model vs the tuned skill.

04 · Audit trail

Every skill version, on file

The lab keeps every skill iteration stored in the dashboard, so you can see which changes made a performance difference.

skilltune.dev — lab
code-review
SKILL.md→ v10
evals/cases.jsonl
schemas/findings.schema.json
versions
v10.mdcases.jsonl
code-review › versions › v10.md
markdown
05 · Export

Take it anywhere, in one click

Export the final version or any version from the lab. Run your tuned skill in any harness with any model.

code-review / SKILL.mdv10 · 92 ✓
pick any version from the lab
also exports to

Your tuned skill is proven to perform better than the base skill.

Runs in every major AI coding tool

An open letter from the founder:
skills are systems, not prompts.

We switch to a bigger model or start fine-tuning when we want better results on our custom workflows. In most cases neither is needed. A well-structured skill can lift any LLM model's performance by up to 20%, with no fine-tuning at all. I tested this with 500+ skills and 20+ models.

In our testing, a model with a tuned skill beats the same base model on both performance and cost. It hallucinates less, burns fewer tokens on silly errors and gets the task done faster with better output quality.

I built skilltune to measure what a skill actually does. The lab writes custom evals from your description, runs them on your target model, and scores every version of the skill. The result is evidence: exactly how much better your model performs with the skill than without it.

Most skills are written once and never tested with any eval cases. Don't trust them blindly. Test your skills so you can use them in your daily workflows with confidence.

Skills are the most underused layer of the AI stack. The skilled models era is here. This is just the start.

CJ Zafir
CEO and Founder
skilltune.dev
How it works

Create a lab-tested skill in four steps

You describe the skill and answer a few questions. The lab creates it, tests it and hands you the lab-tested skill with a full report.

Describe or import

Write what the skill should do, or upload a SKILL.md you already use. Plain words are enough. The lab asks about the rest later.

  • Start from a prompt or an existing SKILL.md
  • No format to learn. Just describe the job.
Input
Create a Skill that writes zero-downtime Postgres migrations — expand first, backfill in batches, never lock writes…
or upload SKILL.mdCreate Skill →

Pick where it runs

Pick the harness and the model your skill has to work in. Then connect the Claude or ChatGPT subscription you already pay for. Tuning runs on your plan, with no separate API bills.

  • Works with Claude Code, Codex, Cursor, opencode and more
  • No API keys. Runs on your subscription.
Setup
harness
Claude CodeCodexCursoropencode
model
GPT-5.6Fable 5Sonnet 5Opus 4.8
account
Claude subscription · connectedConnect ChatGPT
your subscription runs the lab — no separate API bills

The lab fills the gaps

The lab reads your request and asks about anything missing: edge cases, tone, boundaries. Then it writes v1, runs the eval cases and scores it. That score is your baseline.

  • Clarifying questions before any tuning
  • v1 scored right away, your starting line
Draft
1

What exact outcome should this skill produce?

Describe the job to be done and what a successful result looks like.

Type your answer here…
Use defaultVoice✦ Draft answer
2

Who will use this skill, and where does it fit in?

Name the primary user, their expertise, and when they invoke it.

SKILL.md · v1 draftedbaseline · 60

The loop runs until it passes

The lab tests the skill, fixes what fails, and saves the result as a new version. A version stays only if it scores higher than the last one. When a version passes the target score, tuning stops. You get that SKILL.md file, its score graph and the final benchmark.

  • Score graph for every version
  • Final benchmark vs the base model
  • Portable to any harness, any model
Release
+42 vs base model10/10 tests passing
Download SKILL.md · v8
Pricing

One plan. Everything included.

No tiers, no monthly subscription, no add-ons. Every feature is included, with 12 months of regular updates.

Everything in skilltune.dev

  • Create or improve unlimited skills. No rate limits.
  • No credits, no API keys.
  • Runs on your Claude, ChatGPT or opencode subscription.
  • Everything stays on your machine. Full privacy.
  • One-click export to any AI coding tool
  • Works with open source LLM models
  • Access to exclusive skills
  • 12 months of product updates
  • Priority support from the team

One plan, solo or team. Cancel in one click. Every skill you created stays yours.

Early bird · first 500 accounts
$149/year
$199/yearsave 25%
Claim an early-bird seat

After the first 500 accounts, skilltune.dev is $199/year. Early birds keep $149/year for as long as they stay subscribed.
Billed yearly · cancel anytime · price locked while subscribed
FAQ

Questions, answered straight.

The short version of everything people ask before claiming a seat.

Something else? hello@skilltune.dev

01What exactly is a skill?

A SKILL.md file — instructions, examples, guardrails and evals that teach an AI agent to do one job well. It's not just a text prompt but a complete system with scripts, assets and output instructions. With skills you can uplift the base model's performance by up to 20%.

02How does the lab test my skill?

The lab creates eval cases from your description and locks them before tuning starts. Every version runs the same tests and gets a score against the base model. A version stays only if it scores higher than the last one. When a version passes the target score, tuning stops.

03Do I need to write evals myself?

No. The lab drafts the tests from your description; you review and approve them before any tuning starts. Once approved, they're immutable.

04Can I bring a skill I already use?

Yes. Upload your SKILL.md and the lab benchmarks it, then tunes it version by version until we reach the best version.

05Which AI coding tools and models does it work with?

All the major ones. Your tuned skill exports in one click and runs in Claude Code, Codex, Cursor, opencode or any other AI coding tool. Open source LLM models work too.

06Is my work private?

Yes. Your skills, versions and eval results are stored on your machine. Nothing leaves it.

07What happens after the first 500 seats?

The price moves to $199/year. Early birds keep the $149/year rate for as long as they stay subscribed. You can cancel anytime, in one click, but you won't, cuz you'll love it.

08Who owns what I build?

You do. Every version of every skill is yours to export, self-host and ship — including after you cancel. We do not store any of your skills or drafts on our servers.

Create lab-tested skills and use them anywhere.

The end-to-end skill pipeline in one place: create, test and export. Tune your skills with skilltune.

$149/year early-bird · every version on file · runs in Claude Code, Codex, Cursor, opencode