Every listing publishes its measured trigger accuracy, before and after.
See the catalogueThe ones on GitHub are a good idea that fires half the time and breaks on a machine that isn't the author's. We fix the trigger, test in a container, write three worked examples, and publish the measured before and after for both.
Not code. You can clone the original in one click, and we link to it. You're paying for the day someone else spent making it reliable.
Skill imported
on your task…
Drag the folder into your skills directory and you're done. No package manager, no config file, no account, nothing to sign in to. Ask for the thing in plain words and it fires.
description:
helps with code
never fired
The description was written for a human, so Claude never matches it. The script assumes a tool you don't have. Nine hundred lines of SKILL.md eat the context window, and you stop reaching for it.
should fire
should not fire
vs the original
published
Trigger accuracy, before and after, against a written test suite that ships public. It's the one claim on this site that can be checked.
Updates
Re-tests
Replies
Twelve months of updates as Claude Code moves underneath it, re-tested on a clean machine each time, and a person who answers.
$docker run --rm skilltest
fresh image, no local state
installs and loads
Every skill installs into a fresh container with no local state before it ships. The machine a skill was written on is the one place it is guaranteed to work, and it is the one machine you do not have.
Markdown and scripts in a folder. No proprietary format, no runtime, and nothing to uninstall.
Each one lists what it does, what it needs, and the trigger accuracy we measured before and after.
Writes React the way your team already agreed to write it.
Eval suite not run yet
Cuts what Claude reads before it starts, without cutting what it knows.
Eval suite not run yet
Reviews a screen against real UI rules, not taste.
Eval suite not run yet
Search by the problem rather than the tool's name. Every listing says what we changed and what it came from, before you spend anything.
Describe what keeps going wrong. You do not have to know what the skill is called, or that one exists. Filter by category, by price, or by the size of the measured delta.
Rewrote the trigger description
Checked the tools it assumes
Cut SKILL.md to what Claude reads
Added three worked examples
Every product page lists the failures we found and the fixes we made, with a link to the original so you can compare them yourself.
Your rules, not generic ones
Its own eval suite
A private channel to ask in
Nothing in the catalogue fits the way you work? Tell us the task and we will harden one against your own rules, with its own eval suite.
A skill is instructions Claude reads before it starts. A hardened one fires on the prompt you would actually type, and answers with your own rules rather than with general advice.
We pick a job people redo every week and work out what a reliable skill for it has to do. Nothing ships that we cannot test.
Ten fixed points: clean-machine test, rewritten trigger description, token diet, honest error handling, real docs, an eval suite.
A ZIP and a licence key by email. Unzip, drop it in, done. Updates for a year and a person to ask when it breaks.
Finding a skill worth using, checking its licence, fixing the trigger, testing it somewhere clean and writing the docs is a day's work. It is the same day for every buyer, and we only do it once.
SPDX check before any work starts. GPL and non-commercial never enter the catalogue, however good the skill is.
Ten fixed points: the trigger description rewritten, assumed tools checked, SKILL.md cut to what Claude reads, errors that say what went wrong.
A fresh container, no local state, installed from the ZIP a buyer gets. Nothing ships that has only run on our laptop.
Written before the hardening starts, so the before number is honest. It ships public with the skill and you can run it yourself.
Each skill is priced on its own page, for what that one skill saves you. One-off, with no subscription, no seats to count and nothing that renews while you are not looking.
A ZIP and a licence key within a minute of paying. Inside: the skill, our README with three worked examples, and a troubleshooting table written from the ways it actually broke.
Re-tested on a clean machine whenever Claude Code moves underneath it, and reissued to you. No renewal, no second charge.
If we cannot deliver the skill you paid for, you are refunded in full. Every listing publishes its measured trigger accuracy and its worked examples first, so nothing about it is a surprise after the fact.
The six we are asked most. Everything else is in the install guide, or by email.
You can, and for a one-off task you probably should. The cost is not the first draft: it is the trigger description that fires on the words you actually use, the tools it assumes are installed, the context it eats every turn, and the docs. We do that once and publish the measured number so you can tell whether it worked.
A ZIP of the skill and a licence key, by email, within a minute of paying. Inside: the skill, our README with three worked examples, and a troubleshooting table written from the ways it actually broke on a clean machine.
Nothing technical, and we will not pretend otherwise. Claude reads a skill as plain text, so anything on your disk is copyable. What you are buying is that it keeps working: updates when Claude Code changes, and someone to ask when it does not.
A written set of prompts that should fire the skill and prompts that should not, run against both the original and ours. The suite is public and linked from every product page. If we did not measure it, we do not publish a number.
If we cannot deliver what you paid for, yes, in full. Because a skill is plain text you can read the moment it arrives, we do not offer change-of-mind returns on a delivered file. Everything you need to decide is on the listing before you buy: the measured trigger accuracy, the three worked examples, and what the skill needs installed.
We do. Where a skill builds on someone else's published work we say so on its page and ship their licence with it, but the catalogue is our own code, written to one standard and tested the same way.