refactor: prep skills and agents for claude 5gen models

[ryzizub]

Description

Applies the Claude 5-generation context-engineering rules to Wingspan, mirroring the same prep done in vgv-ai-flutter-plugin#152. Net -483 / +156 lines across 20 files, plus 6 new reference files.

Rule	Applied here
Examples → tool design	Stripped <examples> from 6 agent descriptions and rewrote each to route on its own — what it inspects, what it returns, when to reach for it
Repetition → single source	Merged when_to_use into description across 11 skills; extracted the procedures duplicated across skills into skills/shared/references/
Upfront info → progressive disclosure	Demoted run-time detail out of the SKILL.md bodies into references/
Rules → judgment	Dropped trailing ## Important / ## Key Principles blocks that restated the body
Simple specs → rich references	Documented the KEEP / DEMOTE / DELETE rubric in CONTRIBUTING.md

The agent descriptions are the headline. Those <examples> blocks were concatenated into the agent listing on every request — roughly 2,400 tokens per turn, for agents the skills dispatch by name anyway, so the examples bought no routing accuracy.

New shared references: feature-branch.md (brainstorm, plan, debrief) and review-dispatch.md (build, hotfix, review), symlinked per the existing sharing convention. New skill-local references: commit-autonomy.md and ship-gate.md (build), plan-authoring.md (plan), issue-previews.md (debrief).

Skill sizes: build 249→183, plan 219→173, hotfix 166→150, brainstorm 165→142, debrief 156→118, review 136→116.

Type of Change

• New feature (feat)
• Bug fix (fix)
• Code refactor (refactor)
• Documentation (docs)
• CI change (ci)
• Chore (chore)

Notes for reviewers

• One deliberate difference from chore: 0.0.2 #152. Its rubric targets ~150 lines per SKILL.md because those skills are teaching material. Wingspan's are procedural — the steps are the content — so the rubric here sets no line target. The test is per section: does this change what the model does at the point it is read?
• Hard constraints were kept hard. The hotfix 5-file blast radius, the 3-attempt fix cap, and the success-criteria contract all survive untouched; only generic restatement was cut.
• when_to_use was merged rather than deleted. Claude Code appends it to description in the skill listing, so dropping it outright would have lost trigger phrases.
• claude plugin validate, markdownlint, and cspell pass. Frontmatter parses, names match directories, no broken links or symlinks. skills_lint was not run locally (reusable CI workflow) — CI will confirm.

🤖 Generated with Claude Code