The skill description was written keyword-first ("use when the user
says lazy / complains about boilerplate"), so a model-invoked host
only loaded it on prompts that hit those words. Plain coding tasks
(add a date picker, write a dedupe function, build a cache, fix a bug)
were skipped, exactly the over-build cases ponytail is best at.
Tested via a skill-router proxy (3 runs, 12 labeled prompts): recall
on coding tasks was 2/6 before, 6/6 after, with precision unchanged at
6/6 (no false fires on non-coding prompts). Fix is one clause naming
coding tasks explicitly plus a negative clause to keep precision.
Trigger descriptions live in two source-of-truth spots, both updated:
skills/ponytail/SKILL.md and the DESCRIPTIONS map in
scripts/build-openclaw-skills.js (.openclaw regenerated). The store
blurbs in the plugin manifests are listing metadata, not triggers, so
they are left alone.
Emeriko
0e3fd0cfee