Skills for Claude Code · built for production
Not markdown you hope works — pipelines that ran a real business first, then got scanned, signed, eval-tested, and hardened for you. Start with five, free, today.
Try it now — free
Apache-2.0. Runs in your own environment on your own accounts — we never see your keys or data.
And the tool we measured ourselves with. trigger-probe is one Python file, Apache-2.0, no dependencies — point it at a cases file and it tells you whether your own skill actually loads. It produced the report above, including our eight failures. Giving away the measuring tape costs us nothing and makes every number we publish checkable.
Catch hardcoded keys, IPs and personal data before they ship in a skill.
Find dead paths and stale references across your CLAUDE.md, memory and skills.
Capture real network requests from a demo, then build automation on facts.
Keep background jobs running without console windows stealing focus. Windows.
Cross-language OpenSSL "Salted__" AES interop, one reference page.
Why a paid library at all
Free skills are free for a reason: no versioning, no maintenance, and in independent scanning, 36% of shared agent skills carried prompt-injection techniques (Snyk ToxicSkills, Feb 2026). The paid library is the answer to that.
Every release swept for secrets & PII. Best-effort, honest about its limits.
minisign signatures, public key published out-of-band. What you install is what we shipped.
All 18 measured — 360 cases, fresh session each. Trigger 144/180. Restraint 180/180. Eight skills fail our own bar and we publish them. Full report →
Re-verified within 48h of every model update; fixes shipped or status disclosed.
What is in the library
Built from pipelines that ran a real-estate business in production — session arbitration, approval gates, sync watchdogs — genericized and hardened. Every one ships with its own tests, an eval case set, a changelog, and an honest list of what it does not do.
SkillLock pins installed skills to content hashes and reports drift. Skill Eval Harness measures whether a skill fires when it should. Knowledge Intake scores a claim before it enters memory. Secrets Hygiene Sweeper and Config Doctor catch what shipped by accident.
Watchdog Kit restarts a job when it crashes and when it is alive but stuck. Notify Hub delivers the alert with dedup and rate limiting. Go-Live Checklist verifies DNS, ports, health and certificates before launch. Session Arbiter stops concurrent agents corrupting each other.
Approval Gate puts a human yes between what an agent extracted and your database, on an append-only trail. Sync Pipeline mirrors upstream data — retire instead of delete, refuse a truncated “full” sync, never overwrite what a human wrote.
Draft, screen, publish. The drafter builds MLS, portal and social copy from a facts file, so it cannot state a figure nobody typed — and it measures the character budget instead of guessing, names every sentence it left out, and raises the pre-1978 lead-paint disclosure. The screen applies 21 rules over the protected classes plus the source-of-income and age protections states add, and the steering proxies: neighbourhood safety, school-quality claims. Each finding names the class, says why it draws complaints, and rewrites it to describe the property instead of the occupant. Both offline, so copy tied to a real address never leaves the machine.
Both fail their own trigger evals — 5/10 and 3/10. Name either one and it runs; ask in general terms and the model usually writes or judges the copy itself. It can produce something — just without the character budget, the completeness check, the disclosure trigger, or any record of what was checked. We tried two description rewrites aimed at exactly that and measured no improvement from either, so we stopped guessing and published both failures with the disproved hypotheses. We also published a pattern from these two and then retracted it once all 18 skills were measured and a reference skill scored 10/10 while an infrastructure skill scored 6/10 — the retraction sits in the same report as the data that killed it.
We ran every one of these against our own library before shipping it, and it keeps finding things. The eval harness found three bugs in itself — the last one was reading the wrong table column, so three of one skill's ten cases were being scored against the explanation instead of the prompt. Knowledge Intake found two in its own conflict detection. The secrets sweeper turned out to be reporting filesystem paths as credentials, which also hid the genuine path findings underneath them. And a new loadability check caught one paid skill whose frontmatter did not parse at all: a YAML value cannot contain a colon followed by a space, its description said “Not for: …”, and so that skill had been shipping in a state where it would never fire on its own. All of it is written up in the changelogs, because a tool that has never been caught being wrong has probably never been checked.
Pricing
Two caps, both real and both small. When they are gone the list price is what is left — there is no second founding round.
The fastest way to get value out of this
You should not have to become an expert in a file format to get a job off your plate. So we do it with you:
Custom skill inside 5 working days of an agreed spec. 3 this week, 10 total before the price moves — we would rather deliver three properly than sell thirty. Delivered async in writing by default; ask for a live call if you prefer one.
The workflow, not just the packages. One listing-facts.json per listing, committed, plus a drafts/ directory — then your own sentences mention the file and the right skill loads without you thinking about it. The quickstart ships with the library, and says plainly that this part is a recommendation we have not measured, because it changes what a person types rather than how the model chooses.
Questions we would ask
If an answer here would put you off buying, it is better that it does so now.
Because free ones have no version, no maintenance, and nobody accountable when a model release changes their behaviour. Every release here is swept for secrets and PII, signed with minisign, ships an eval case set with published numbers, and gets re-verified within 48 hours of a model release — with fixes shipped or the status disclosed in the changelog. If none of that is worth $199 a year to you, the five free skills are genuinely free and we would rather you used those.
Some do not, and we publish the numbers for all 18. Trigger 144/180: ten skills clear our 90% bar and eight do not, the worst at 3/10. About seven in ten misses are the model doing the job itself instead of loading anything — naming the skill works. Every scorecard lists its failing cases and the fixes we tried that did not help, and the library report includes a pattern we published and then retracted when more data killed it.
Yes, and you should. Five skills are Apache-2.0 and install with one command — the secrets sweeper, config doctor, learn-by-demo and silent runner are cut-down versions of paid ones, so you are evaluating the real engineering, not a landing page. The Done-with-you engagement also agrees scope in a public thread before any invoice.
No. Skills run in your environment on your accounts, and every engine in the library is Python standard library only with no network-capable module in the code path — verified by test, not by promise. We have no server for your data to reach.
7 days on monthly, 14 days on standard annual, and 30 days on Founding Annual and Founding Lifetime — the longer window is a launch-promotion exception, not the standing policy. This does not limit your statutory rights under applicable consumer law. A chargeback costs us more than a refund does, so asking is always the faster route.
Repository access ends and the files you already installed stay on your disk — they are files. You lose updates, re-verification after model releases, and new skills. Founding Lifetime does not renew and does not end. Nothing phones home to check.
No, and we are not affiliated with, sponsored by, or endorsed by Anthropic. We build on the open Agent Skills format, which is why the same packages work across agents that read it. "Claude" and "Claude Code" appear here only to describe compatibility.
A small operation, funded by revenue rather than a runway — which is the honest version of a survival claim. What we can show instead of a promise: the changelogs, the signed manifests, the eval numbers including the failures, and the free repo you can fork today so that nothing here can be taken away from you.