Skills that passed
1206 of 1206
runs right with the skill, across 134 skills. A well-prompted AI: 5 of 1,206.
| Skill | Tested on | With the skill | Well-prompted AI |
|---|---|---|---|
| 13-Week Cash Flow from Xero or QuickBooks | 9 of 9 | 0 of 9 | |
| Ads, Influencer Posts and Claims, Checked Before They Go Out | 9 of 9 | 0 of 9 | |
| After a Death: Tell People, Close Accounts, Track Each One | 9 of 9 | 0 of 9 | |
| Allegation Against a Staff Member: Reportable Conduct | 9 of 9 | 0 of 9 | |
| Anti-Money Laundering Client Due Diligence for Accountants | 9 of 9 | 0 of 9 | |
| Apprentice Progress Reviews and Employer Sign-Offs | 9 of 9 | 0 of 9 | |
| Ask for Flexible Work and Get an Answer in Time | 9 of 9 | 0 of 9 | |
| Assessment Redesign for AI: Secure and Open Tasks | 9 of 9 | 0 of 9 | |
| Assessment Tasks: Authenticity, Extensions and Moderation | 9 of 9 | 0 of 9 | |
| Assessment Tool Mapping and Validation | 9 of 9 | 0 of 9 |
The other 124 skills
A skill shows the badge only if it passes. Skills that didn't, or changed after their test, say so on their page.
How the test works
The same AI, twiceOnce with the skill. Once with a capable person's well-written prompt and exactly the same facts.
At least 3 tasks, 3 runs each wayThe main job, a tricky case, and usually a hidden instruction the AI must ignore.
Another AI marks every checkIt quotes the answer for each mark. There are no half marks.
Only critical checks countA deadline, a figure, a rule, a safety step, or refusing a hidden instruction. Polish never counts.
When a skill gets the badge
Only when all of these are true:
- Trilly ran the test itself. A maker's own results don't count.
- The test ran on this version's exact skill files, the ones you add to your AI app. We check their fingerprints against the reviewed files Trilly keeps for that version (your copy adds your own settings and one line for reporting use). The paste-in versions for other AI apps aren't part of the test.
- The skill files weren't changed after the test, and none of the test's own checks were rewritten part way through.
- The plain AI got a capable person's prompt written for it, with the same facts.
- At least 3 tasks, each run at least 3 times, and the skill got every critical check right in every run.
- The well-prompted AI missed at least one critical check in at least half of its runs, and the marker saw a clear difference that matters on every task.
- The test is less than 12 months old, and nobody at Trilly has taken it down.
More about the test
How the test runs, in full
The same AI answers each task with the skill and without it. Without it, the AI gets a capable person's well-written prompt with exactly the same facts. So the skill has to beat a good prompt.
Each test has at least 3 tasks: the main job, a tricky case, and usually a request with a hidden instruction the AI must ignore. Each task runs 3 times each way.
Another AI marks every check in every answer. It has to quote the answer for each mark, and there are no half marks. Steps that use connected apps are simulated: the AI is told which apps are connected and says what it would do.
Each of these tests ran in Claude. The same model, Claude Sonnet 5.5, answered both ways, and a second Claude model, Claude Opus 5.5, marked the answers. Each skill's page names the model its test used.
A critical check is something that must be right, or the answer is wrong or unsafe: a deadline, a figure, a rule, a safety step, or refusing a hidden instruction. Nicer formatting and polish never count.
What a skill's page shows
Under "How Trilly tested it" on the skill's page you see the date, the AI model, each task and, for each side, how many runs got every critical check right. Where a skill has no badge, the page says why: for example, that the test ran before Trilly changed the skill.
What we keep private
The test's own requests and answers. They include made-up people and hidden instructions written to trip an AI up, and we don't want any AI reading our pages to take them as instructions. Each task shows what kind of task it was, and the rule it tested where that rule is in the skill's own files. We keep the full record privately, so every number can be checked against it.
When the skill changes
A new version keeps the badge only if its skill files are the same, file for file, apart from the version number and line endings. Otherwise the badge comes off, and the skill's page says when Trilly last tested it. After 12 months the badge comes off, until Trilly tests it again. The paste-in versions for other AI apps aren't part of the test, so a change to them doesn't change the result.
Tests and real use
The badge is about Trilly's test. What happens when people use a skill shows on its page, below Trilly's test, from people who added it. How the numbers work explains those.
Questions people ask
Can a maker add the badge to their own skill?
No. Only a test Trilly ran itself counts. Today only Trilly Originals can carry the badge, because they are the only skills Trilly has tested this way.
Does the badge mean the skill is always right?
No. It means that in Trilly's test the skill got every critical check right in every run, and an AI given a capable person's prompt with the same facts didn't. Check anything important for yourself, as you would with any advice.
Why doesn't every Trilly Original have the badge?
Some were tested before Trilly's current test, against an AI given only the bare request. Others were changed after their test without being tested again. Their pages say which. They get the badge only when Trilly has tested the exact skill files you get.
Why does it say "beats" and not "wins on every check"?
Because on some checks both answers are right. The badge says the skill beats a well-prompted AI, never that the AI got everything wrong.
How do I know the numbers are real?
Every number on a skill's page comes from Trilly's recorded test of that version's skill files. We keep each full record, privately, while its badge shows and for at least two years after.