Trilly's own skills that passed its test against a well-prompted AI
The scoreboard. From our own records.
This quarter, since 1 October.
One version, every check. The copy you add is the copy we checked.
Each version gets its own fingerprint. If its instructions change, the checks run again, and the test badge comes off until Trilly tests it again.
When Trilly adds a new Claude model to its list, it can run the safety test again with that model, and a skill that passes says on its page when and with which model. That line says what was run, nothing more.
What we check on every version
Hidden instructionsText you can't see, odd characters, attempts to take over your AI.
Sending your information awayBlocked, to websites the maker hasn't told us about.
Safe files onlyAllowed types and sizes. Nothing that installs software.
Leaked passwords and keysNo private details left inside the skill.
Does what it saysNo fake reviews, made-up numbers or promises it can't keep.
A behaviour testA safe test chat with pretend tools, watched.
An AI review of every text fileNever passed with any part unread.
A full review for risky skillsMoney, betting, health, legal and account-changing skills.
Trilly approves big decisionsTrilly approves each version about money, betting, health, the law or your accounts.
Help to learn, never the work done for themRead the rule for study skills
Read what we check, in full
Every skill on Trilly is checked before anyone can add it, and checked again every time it changes. Skills that deal with money, betting, health, the law or your accounts also get our full review, and Trilly approves each version before it goes live. If a skill breaks our rules, we take it down and say so in public.
Hidden instructions. We look for text you can't see, odd characters and instructions that try to make your AI ignore you or break its own rules.
Sending your information away. We block skills that try to send what you share to websites the maker hasn't told us about.
Safe files only. Skill packages can only hold allowed file types and sizes. Nothing that installs software on your computer without you knowing.
Leaked passwords and keys. We check that makers haven't left passwords, keys or other private details inside the skill.
Does what it says. The listing must match what the skill does. No fake reviews, made-up numbers or promises it can't keep.
A full review for risky skills. Money, betting, health, legal and account-changing skills are checked against our full review list before they go live, and carry a plain warning.
A behaviour test. We run each skill in a safe test chat with pretend tools, and watch what it tries to do. If a check can't run, the skill waits for our full review instead.
An AI review of every text file. An AI reads every text file in full. If a skill is longer than 600,000 characters, the rest is read in our full review. If a check can't run, the skill waits for our full review instead. A skill is never passed with any part unread.
Trilly approves big decisions. Each version of a skill about money, betting, health, the law or your accounts needs Trilly's approval before it goes live.
Each version of a skill gets its own fingerprint, so the copy you add is exactly the copy we checked. Full rules: what skills may not do.
After you add a skill.
Safety and what it can reach. Before you add a skill, its page lists the websites, data and accounts it can reach.
Updates wait for your yes. A linked skill gets an update by itself when it needs nothing new. A downloaded copy stays as it is until you download the new version. If an update needs more access than your version, a linked skill waits for your yes, and you keep the version you agreed to until then. When a skill updates.
Signed copies. Downloaded copies are signed, so anyone can check one against what Trilly sent (paste-in versions can't be signed). When a copy reports use, it includes a fingerprint of its instructions, and one that doesn't match is left out of the public numbers. A linked skill always loads the checked version, straight from Trilly.
Help to learn, never the work done for them. No skill on Trilly may write, rewrite, paraphrase or complete work a student will hand in as their own, sit a test for them, or help hide AI use from a teacher. Skills that explain, quiz or give feedback on a student's own writing are welcome. We check every new version for this, and remove any skill that breaks the rule. Read the rule.
Your feedback reaches the maker, and you hear when it's fixed.
ExampleFeedback goes to the maker only with your OK.
Someone's AI, sent with their OK
"The invoice total skipped tax on one line."
The maker sees a theme
Tax on one line
7 reports, personal details taken out
A new version, checked again
Version 1.3
Fixes 7 problems people reported
The 7 people are told
"Version 1.3 fixes what you told its maker."
What the badges mean
What the badges mean
Checked by TrillyThis version passed our checks. Every listed skill has it.
Beats a well-prompted AITrilly tested this version's skill files itself against an AI given a capable person's prompt with the same facts, and the skill got every critical check right in every run. A maker can't add it. How we test
Trilly OriginalMade by the Trilly team, with AI, and checked like every other skill before it's listed. Its use numbers come only from real use, the same as anyone's. About Trilly Originals
Real person · ID confirmedWe've confirmed the maker is who they say they are.
OfficialMade by the company whose product the skill works with, and that company has proven it owns its web address.
ApprovedMade by someone else, but the company has said it is happy with it.
Helps you learn: you do the workTrilly's review (never the skill's maker) marked this version as one for learning: it explains, quizzes or gives feedback on your own work, and never does the work you hand in. A maker can't add it. Each new version is decided again, so a skill can lose it. Read the rule
Report a problem, and we look at it.
Use "Report" on any skill, or contact us. A report never takes a skill down by itself. We remove an unsafe or rule-breaking skill at once, tell the people who added it while signed in, and list it publicly.
Questions people ask
Can a skill steal my information?
A skill is a set of instructions. It works with what you share in your chat and, if you've connected apps or tools to your AI app, with those too. A skill that changes anything in your apps must ask you first. We block skills that try to send your information to other websites, hide instructions, or trick your AI into ignoring you. If you ever see a skill do something odd, report it.
Who writes the reviews?
Only people who added and used the skill. Each review says how long the person used it, and makers can reply but can't delete reviews.
