Agent Skills
eworthing/agent-skills
Skills reutilizáveis para agentes IA de codificação com foco em ecossistema Apple
Ver no GitHubBiblioteca de skills reutilizáveis para agentes de IA de codificação (Claude Code, Codex, Gemini CLI, Copilot CLI). Fornece referências especializadas em desenvolvimento Apple (SwiftUI/UIKit), segurança iOS, padrões de persistência, design tokens e componentes nativos para acelerar implementações em múltiplas plataformas.
Instalação
npx convez-ai eworthing/agent-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Desenvolvimento cross-platform Apple (iOS, macOS, tvOS)
- Implementação de segurança em aplicações iOS
- Padrões SwiftUI e otimização de arquitetura
- Integração de design tokens e UX nativa
- Automação de linting e divisão de arquivos Swift
Veredito

Excelente repositório de skills especializadas para agentes IA que trabalham com código Apple. Cobertura profunda em SwiftUI, segurança e padrões nativos que seria tedioso o agente redescobrir sozinho. Recomendado para equipes que dependem de AI coding assistants em projetos iOS/macOS.
README
A tradução para o português deste README ainda está na fila.
Agent Skills
Reusable skills for AI coding agents — Claude Code, Codex CLI, opencode, Gemini CLI, and Copilot CLI.
Single source of truth lives here; agent CLIs consume each skill through a symlink (see Deployment).
Skills
Apple platform — SwiftUI / UIKit / Foundation
- apple-multiplatform — Cross-platform Apple SwiftUI compatibility reference for iOS, iPadOS, macOS, Mac Catalyst, and tvOS. Platform-conditional code,
#if os()vs#if canImport(), gatingeditMode/ drag-receiving / haptics for tvOS,TabView.pagevs.automatic, Mac Catalyst sidebar/window defaults,XCUICoordinate/NSToolbarUI-test divergence. - apple-tvos — Definitive tvOS reference for SwiftUI on Apple TV: focus engine, accessibility deltas, and design-regression checks. Settle delays,
.focusable()container blocking,.onExitCommandMenu dismissal, destructiveconfirmationDialogdefault focus, modal focus containment, glass-on-glass on tvOS. - ios-security-hardening — Input-validation and file-handling safeguards for untrusted data: path-traversal prevention, URL scheme/domain allowlisting, multi-source image-reference resolution, CSV/JSON sanitization, AI prompt sanitization, sandbox directory usage, iOS Data Protection levels.
- swift-file-splitting — Splits oversized Swift files into smaller units while preserving visibility and build correctness. SwiftLint
file_lengthviolations, type/extension extraction. - swift-linting — Resolves repository-specific SwiftFormat / SwiftLint rule violations.
function_body_length/type_body_length/file_length/cyclomatic_complexity/line_lengthfixes, justified// swiftlint:disable:nextwith rationale, reconciling SwiftFormat output with hand-formatted code. - swiftdata-persistence — SwiftData patterns and gotchas for
@Model,ModelContext,ModelContainer,FetchDescriptor, migrations, cascade-delete relationships. Bundled seed data, "data not showing" / "stale entity" / "images-show-placeholder-after-upgrade" diagnostics, auto-saving on a timer. - swiftui-design-tokens — Applies project design tokens for colors, spacing, typography, motion, and button styling in SwiftUI on iOS, macOS, and tvOS. Spring/timed motion tokens, Reduce Motion alternatives, replacing hardcoded values, macOS form styling, modal frame sizing.
- swiftui-drag-drop — SwiftUI drag-and-drop architecture for iOS, iPadOS, macOS.
DropDelegatevs.onDrop, drop-priority routing, multi-provider payload extraction, NSItemProvider lifecycle, Chrome image drag (public.tiff/public.html/public.url), Button drop-attachment pitfalls. - swiftui-file-export — SwiftUI file export with the modern
TransferableAPI on iOS 16+ / macOS 13+.fileExporterroot-placement rules,ShareLinkvsfileExporter, sandbox compliance, macOS entitlements, menu-barCommandsintegration, "fileExporter silent failure on macOS" diagnosis. - swiftui-native-ux — Native iPhone/iPad SwiftUI design intelligence.
TabViewvsNavigationStackvsNavigationSplitView, empty/detail pane states, iPhone-to-iPad adaptation, sheet vs inspector vs sidebar, Liquid Glass placement, anti-web-smell critique, Dynamic Type / VoiceOver / Reduce Transparency review. - xctest-ui-testing — XCTest UI automation for iOS, macOS, tvOS. Typed-enum accessibility identifiers, wait-for-element, drag-and-drop testing, sheet/alert coverage, macOS window-pinning,
.xctestrun-based selective execution (-only-testing, list/range/glob/match/class/id), zero-test detection,-retry-tests-on-failure.
Cross-language tooling
- bash-macos — Keeps shell scripts portable across macOS (Bash 3.2, BSD userland) and Linux (Bash 4+, GNU coreutils). Debugs "command not found" / "invalid option" / "mapfile: command not found", GNU-vs-BSD
sed/grep/date/readlinkissues, snake_case verb-first renames. - doc-standardization — Standardizes documentation naming, organization, and cross-references in Markdown projects. Enforces
[domain]-[feature]-[type]-[status].mdfilename pattern, valid internal links, ordered index files, code-to-doc identifier alignment.
Workflow / meta
- contest-refactor — Autonomous Actor-Critic refactoring loop. Aggressively refactors the workspace to a 9.5+ standard using a strict ICA-grounded architectural rubric (deletion test, two-adapter rule, depth-as-leverage). Invoke via
/contest-refactor. - peer-plan-review — Sends an implementation plan to another AI agent (Codex, Claude Code, Copilot, opencode, Antigravity
agy, or Gemini CLI) for iterative review, revises, and re-submits until approval or round limit. Provider adapters, session resume, process-tree timeout kill. Usage:/peer-plan-review <codex|claude|copilot|opencode|agy|gemini> [model] [effort]. (agyis experimental — not guaranteed read-only; Gemini CLI is EOL 2026-06-18, enterprise-only.)
Experimental (not symlinked by default)
- quorum-review — Multi-provider consensus review system (v3.1). Anonymous quorum reviews for plans/specs/diffs with canonical issue IDs, conservative merges, independent verifier. Reviewers: claude, codex, copilot, agy (Antigravity, experimental), gemini. v3.1 refactor split the orchestrator into a
quorum/package and vendored shared infrastructure from/common/. Source in-repo only — install manually if you want to try it.
Deployment
Each agent CLI looks for skills in its own per-user directory. Best practice: symlink the repo skill into each directory — never copy.
| Agent | Discovery path |
|---|---|
| Claude Code | ~/.claude/skills/<skill> |
| Codex CLI | ~/.codex/skills/<skill> |
| opencode | ~/.config/opencode/skills/<skill> |
| Gemini CLI | ~/.agents/skills/<skill> (shared community location) |
| Gemini Antigravity CLI | ~/.gemini/antigravity-cli/skills/<skill> |
Note on
~/.agents/skills/: this directory is also used by external/community skills (swiftui-expert-skill, swift-concurrency, etc.). The install loop below skips any existing entry that isn't already a symlink to this repo — community skills are never clobbered. Gemini CLI also offersgemini skills link <path>as a first-party alternative.
Why symlink, not copy
- One source of truth — edit the file in this repo and every agent picks it up immediately on the next session.
- No drift between agents.
git pullupdates all three CLIs simultaneously.- Easy to opt out:
rm ~/.claude/skills/<skill>disables for one agent without touching the repo.
Install one skill
SRC="$(pwd)" # run from this repo's root
SKILL=peer-plan-review # change per skill
ln -s "$SRC/$SKILL" "$HOME/.claude/skills/$SKILL"
ln -s "$SRC/$SKILL" "$HOME/.codex/skills/$SKILL"
ln -s "$SRC/$SKILL" "$HOME/.config/opencode/skills/$SKILL"
ln -s "$SRC/$SKILL" "$HOME/.agents/skills/$SKILL" # Gemini CLI
ln -s "$SRC/$SKILL" "$HOME/.gemini/antigravity-cli/skills/$SKILL" # Gemini Antigravity CLI
Install every published skill into all four agents
quorum-review is excluded — still under test. Add it manually if you want to try it.
SRC="$(pwd)"
SKILLS=(
apple-multiplatform apple-tvos bash-macos contest-refactor
doc-standardization ios-security-hardening peer-plan-review
swift-file-splitting swift-linting swiftdata-persistence
swiftui-design-tokens swiftui-drag-drop swiftui-file-export
swiftui-native-ux xctest-ui-testing
)
DESTS=(
"$HOME/.claude/skills"
"$HOME/.codex/skills"
"$HOME/.config/opencode/skills"
"$HOME/.agents/skills" # Gemini CLI (shared with community skills)
"$HOME/.gemini/antigravity-cli/skills" # Gemini Antigravity CLI
)
for dest in "${DESTS[@]}"; do
mkdir -p "$dest"
for skill in "${SKILLS[@]}"; do
target="$dest/$skill"
if [ -e "$target" ] && [ ! -L "$target" ]; then
echo "skip: $target exists and is not a symlink" >&2
continue
fi
if [ -L "$target" ]; then
current="$(readlink "$target")"
expected="$SRC/$skill"
if [ "$current" != "$expected" ]; then
echo "skip: $target -> $current (different source; not overwriting)" >&2
continue
fi
fi
ln -snf "$SRC/$skill" "$target"
done
done
The loop is idempotent and safe in shared dirs:
- Real files at the target path are skipped with a warning.
- Symlinks already pointing at this repo are refreshed via
ln -snf. - Symlinks pointing at a different source (e.g. community skills in
~/.agents/skills/) are skipped, not clobbered.
Verify
for dest in ~/.claude/skills ~/.codex/skills ~/.config/opencode/skills ~/.agents/skills ~/.gemini/antigravity-cli/skills; do
echo "== $dest =="
ls -la "$dest" | grep -- '-> '"$(pwd)" || echo "(none from this repo)"
done
Broken symlinks (target moved or renamed) appear as red entries under ls -la on most terminals; fix with rm + re-ln.
Codex YAML alternative
Codex also accepts agents/openai.yaml-style declarations if you prefer YAML-configured agents over symlink discovery.
Repo Conventions
- One directory per skill at the repo root.
- Each skill:
SKILL.md(YAML frontmatter + body), optionalreferences/(progressive disclosure), optionalscripts/(helper executables), andEVAL.md. - Skill descriptions lead with capabilities and contain a "Use when…" trigger phrase (required by
scripts/eval-skill.py). - Shell scripts target portable Bash (macOS 3.2 + Linux 4+) — see
bash-macos. - Each skill ships an
EVAL.mdscored against the multi-framework rubric in the localskill-evaluatorplugin (.claude/skills/skill-evaluator-1.0.0/). Target ≥ 90/100 before publishing. - Commit style:
feat(<skill>): <change> (<old>→<new> EVAL)when an eval score shifts. Otherwise standard Conventional Commits.
License
MIT