Pick by task.Check before you load.
A short routine for choosing an agent skill: match it to your task, read the pinned source, check the evidence that exists, and prove it on something small before real work.
Choose by task fit
Describe your task in one sentence, with the outcome you need. With Edge set up, your agent searches the public skill catalog and returns ranked matches, each with its description, a pinned source and a security state.
Read the top few descriptions
Prefer the skill whose described triggers match your task. An install count is dated catalog metadata, not proof of quality or safety.
Watch the honest signals
Edge says how many of the top results it considers a good fit, and when it reordered the catalog. If it reports no good fits, treat the matches as starting points, not answers.
Compare dated examples
The task answers show real find_skill results with their dates, pinned sources and runners-up, so you can see what a match looks like before you run your own.
Read the pinned source
A skill is a set of instructions your agent will follow. Edge links the pinned SKILL.md at an exact revision, so you can read precisely what would load.
Read what it tells your agent to do
Look for network calls and which hosts they go to, package installs, and any request for keys, tokens or local files. Anything that sends your data somewhere should have a reason you accept.
Check the repository
Publisher, licence, history. Careful No detectable licence, or an anonymous repository created recently, is a reason to slow down, not proof of harm.
Check the evidence
Edge reports a security state for the exact package bundle and withholds skills with a high or critical finding or two failing scanners; a skill with too little current evidence is shown with a security note. "Reviewed" means recorded Edge-run static checks are current and tied to this revision. It is not a safety guarantee.
Look for a scan evidence page
Some packages have dated, revision-bound scan evidence: the scanner, its version, the scan date, the scope and the counted findings. A pass means no findings were counted within that scanner's stated scope. Static scans check code, not runtime behaviour. The methodology states the limits.
Know what a match is not
A search result, even a first result, is not a security pass and not a task outcome. How Edge checks skills explains the states and what happens at load time.
Prove it on a small task
Start small
Load the skill for one low-stakes task. Watch what it does: files it touches, commands it runs, calls it makes.
Close the loop
If the skill helped, rate the result with rate_skill. If something was wrong, use report_issue. Your feedback gives Edge a concrete case to review.
For developers: full reference and exact inputs
Everything on this page maps to fields you can inspect yourself.
What find_skill returns
Each candidate comes with its source (publisher/repository), skill name, catalog description, install count where the catalog has one, and a pinned source URL that includes the exact revision hash. The response also says how many of the top results Edge considers a good fit for the query, and marks any candidate Edge moved up from the catalog index's own order. A security state line applies to the exact-bundle results; unknown, stale and blocked packages are withheld.
What an evidence page records
A scan evidence page is bound to one package hash and one commit. It records the scanner provider and version, scan date, scope, status, counted findings and raw findings where supplied, plus the licence status Edge could verify. The state definitions are published as data at security-states.json, and the records download is on the dataset page. Every HTML skill page in the evidence catalog has a Markdown twin served with a canonical link header.
What happens at load
Loading through Edge is temporary and pinned to the revision you saw. The load policy is checked again at that moment, so a package whose state changed since a dated page or match can be refused. See Safety for the state list and Setup for adding Edge to your agent.