Skip to content

refactor: modernize plugin skills and agents for claude 5 - #277

Draft
FrkAk wants to merge 1 commit into
mainfrom
refactor/plugin-skills-claude5-t1
Draft

refactor: modernize plugin skills and agents for claude 5#277
FrkAk wants to merge 1 commit into
mainfrom
refactor/plugin-skills-claude5-t1

Conversation

@FrkAk

@FrkAk FrkAk commented Jul 30, 2026

Copy link
Copy Markdown
Owner

Summary

Task Reference: [PYZ-379]

Rewrites the plugin instruction surface for Claude 5 generation models per Anthropic's current authoring guidance: judgment-based prose over ALL-CAPS rule lists, lazy reference loading over eager includes, and harness-enforced tool permissions over prose permission sections.

  • Removes the three @-include blocks that force-loaded the full reference doctrine into composer and the decompose agents; the existing read-when tables remain.
  • Removes the trailing ALWAYS/NEVER recap sections, the rationalization table, stakes preambles, duplicated persona/tone boilerplate, and mid-session re-read scaffolding (ALL-CAPS emphasis: 111 to 6; eager includes: 14 to 4, keeping only the composer phase-rules).
  • Moves tool permissions from prose sections into agent frontmatter tools: allowlists (both mcp__piyaz and plugin-namespaced spellings).
  • Adds a single Hard rules block to the piyaz skill stating the load-bearing constraints once in the active path, including the done-transition rule with its no-questions-pressure nuance.
  • Brand copies regenerated via sync:plugins; extract pin refreshed for the lifecycle reference.

Verified with paired behavioral evals on Opus 5 against the previous skill set (same scenarios, same judges, multiple runs per scenario): zero guardrail violations on the new set versus two on the old, no scenario regressed beyond single-run variance, four improved.

Type of change

  • Refactor / cleanup

Testing

  • Linting passes (bun run lint)
  • Typecheck passes (bun run typecheck)
  • bun run check:plugins passes (brand sync clean)

Notes for reviewer

Deletion-heavy by design; no procedure content was rewritten. The one addition (Hard rules block) exists because the first eval round showed the done-guardrail regressing when the recaps were deleted without an explicit single statement in the skill's active path.

Docs impact

Generated skills/agents catalog syncs automatically via the Docs Sync workflow; no hand-edited pages affected.

Compliance impact

  • none

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant