skills/ponytail/SKILL.md:15 (first paragraph):
End your reply with one or two lines: what you skipped or did not check, and any risk the user must know.
Two failure modes in practice. Evidence from my own sessions since 5.1.0 (pi agent, 28 turn-final replies over two days, 18 of which end with a disclosure):
Scope. The sentence has no coding qualifier, and the v5 frontmatter dropped v4's "Do NOT use for non-coding requests (general knowledge, prose, translation, summaries, recipes)". Once the skill is active for a coding task, research/Q&A/meta replies in the same session also grow a closing disclosure: 9 of my 17 non-coding replies ended with one — a version-diff explanation, a benchmark-numbers summary, an issue-submission confirmation. The disclosures there are manufactured: "did not pull the per-task benchmark tables" on a reply that only relayed the README. This is the same over-application #978 reported against v4's Persistence block; the v5 rewrite removed that block, but also removed the carve-out, and this sentence is stronger than the old wording was.
Repetition. The rule is per-reply, but what it asks about is per-task state. Debugging a game plugin across four consecutive turns, every reply ended "not verified: did not launch the game" — informative the first time, boilerplate the fourth.
The first disclosure in a task is genuinely useful (in mine: a sideloading cert permanently left in TrustedPeople, an attribution flagged as inference), so this is scoping, not removal. Suggested replacement:
End a coding reply with one or two lines when they carry real news: what you skipped or did not check, and any risk the user must know. Nothing new since your last disclosure, or the turn is not coding: end without them.
skills/ponytail/SKILL.md:15(first paragraph):Two failure modes in practice. Evidence from my own sessions since 5.1.0 (pi agent, 28 turn-final replies over two days, 18 of which end with a disclosure):
Scope. The sentence has no coding qualifier, and the v5 frontmatter dropped v4's "Do NOT use for non-coding requests (general knowledge, prose, translation, summaries, recipes)". Once the skill is active for a coding task, research/Q&A/meta replies in the same session also grow a closing disclosure: 9 of my 17 non-coding replies ended with one — a version-diff explanation, a benchmark-numbers summary, an issue-submission confirmation. The disclosures there are manufactured: "did not pull the per-task benchmark tables" on a reply that only relayed the README. This is the same over-application #978 reported against v4's Persistence block; the v5 rewrite removed that block, but also removed the carve-out, and this sentence is stronger than the old wording was.
Repetition. The rule is per-reply, but what it asks about is per-task state. Debugging a game plugin across four consecutive turns, every reply ended "not verified: did not launch the game" — informative the first time, boilerplate the fourth.
The first disclosure in a task is genuinely useful (in mine: a sideloading cert permanently left in TrustedPeople, an attribution flagged as inference), so this is scoping, not removal. Suggested replacement: