Stealth Engine — Agent Skills

A collection of agent skills we use daily in the studio — reusable, on-demand context and instructions for AI coding agents (Claude Code, Cursor, and other skills.sh-compatible agents).

Each skill is a folder under skills/ containing a SKILL.md: YAML frontmatter (name + a trigger description) and a markdown body the agent loads only when the skill is relevant.

Install

Using the skills CLI (no install needed):

# Interactive — pick which skills (and which agents) to install:
npx skills add stealth-engine/skills

# Browse what's in here without installing:
npx skills add stealth-engine/skills --list

# Install a specific skill:
npx skills add stealth-engine/skills --skill safari26-liquid-glass

# Install everything, globally (available in every project):
npx skills add stealth-engine/skills --all -g

Manage installed skills with npx skills list, npx skills update, and npx skills remove.

Skills

Skill	What it's for
safari26-liquid-glass	How iOS 26 / iPadOS 26 Safari's "Liquid Glass" status & address bars interact with web content — the viewport/keyboard facts, why position:fixed breaks the bleed, the keyboard layout bug (WebKit #297779), and case-dependent fixes for immersive/edge-to-edge designs, custom drawers, and themeable backgrounds.
agents-md-setup	Set up a project's agent instructions as one source of truth: AGENTS.md as the real file with CLAUDE.md (and other tools' files) symlinked to it. Bootstraps a repo, converts a standalone CLAUDE.md, or fixes duplicated/drifting instruction files.
skill-publishing	How to author and publish agent skills for the npx skills ecosystem — the SKILL.md format, the description-as-trigger rule, multi-skill repo layout, the skills CLI, publishing to GitHub with SEO, and public vs private repos.
nextjs-locale-standalone	Add locale-prefixed i18n routing to a single Next.js App Router site — a middleware/proxy that redirects / to /<locale>/… by the toggle's last choice (NEXT_LOCALE cookie) then Accept-Language, a [locale] layout + LocaleProvider/hooks, and a LocaleToggle. Ships copy-paste templates.
nextjs-locale-monorepo	The same locale routing factored into a shared workspace package (Turborepo / pnpm) every app consumes — detection + a middleware factory + a use client provider with a dual ESM/CJS build, a shared supportedLanguages config, and thin per-app wiring. Ships the package + app templates.
code-complexity-stats-pr	A GitHub Actions workflow that runs scc on every PR and posts code-size + a COCOMO "cost to build" estimate as a self-updating sticky comment. Covers fork-PR token limits, runner arch, version pinning, comment pagination, and the 65k comment cap. Ships the workflow template.
conventional-commits	The Conventional Commits / "semantic commits" format — type(scope): description, the type→semver mapping, breaking changes, monorepo scopes, and squash-merge PR titles — and why non-conventional messages silently skip releases. The foundation for automated versioning + changelogs.
semantic-release-automation	Automate versioning, changelog, tags, GitHub Releases, and npm publishing from Conventional Commits with semantic-release — the plugin pipeline, single-package vs monorepo (per-package tags + paths-filter matrix) flavors, the Actions workflow + no-loop guard, and pooled releases. Ships .releaserc + workflow templates.
production-release-gating	Stop every merge to main from shipping to prod — deploy only on a real release. Three named patterns: Release-Triggered Deploy (deploy on GitHub Release published, for GKE/self-hosted/CLI), Promotion Branch (fast-forward a production branch the platform watches — most portable), and Build-Skip Gate (the Vercel ignoreCommand script, exit 1 = build, 0 = skip). Works with per-merge and pooled releases; branch-aware previews and monorepo dependency-release handling. Ships all three templates.
git-trunk-branch-and-pr-automation	Trunk-based branch naming (feature/fix/hotfix + AI-agent prefixes claude/cursor/codex/…) and squash-merge where the PR title becomes the Conventional Commit (body = concatenated commits). Ships branch-name validation + PR-title normaliser workflows and the shared script, plus the GitHub squash settings to configure.
autonomous-pr-driver	Drive a PR to merge-ready: open it (or attach to an existing one), watch checks, and run the bot-review loop — triage findings, fix the valid, reject the invalid with a comment (@-mentioning bots that learn), push, repeat until green, then ping a human to merge. Covers dedup-by-finding-ID, verify-before-trust, conflicting-bot adjudication, fork-PR safety, and convergence. Ships a triage playbook + dated bot snapshot.
pooled-release	Cut fewer, batched releases (a "release train") instead of one per merge — trigger semantic-release on demand (workflow_dispatch) or on a cadence (schedule) rather than on every push to main, or via prerelease channels (beta/next). Builds on semantic-release-automation (only the trigger changes); covers why it doesn't slow dev and when a release branch is (rarely) worth it. Ships the workflow + channel-config templates.
resolve-merge-conflicts	Resolve git merge/rebase conflicts non-destructively (preserve both sides' intent) and escalate when a safe resolution isn't clear. Covers rebase-vs-merge, the rebase ours/theirs inversion, diff3/rerere, lockfile/rename/semantic conflicts, abort/reflog escape hatches, verify-after, and fork-PR hand-off. Pairs with autonomous-pr-driver (behind-base / mergeable=CONFLICTING). Ships a git/gh playbook.
dual-agent-review	Review a diff with two+ independent reviewers in parallel — a Sonnet subagent and the CodeRabbit CLI — running as background subagents that surface findings to the orchestrator, which dedupes across reviewers, triages valid/invalid (verify-before-trust; adjudicates disagreements), fixes directly or fans out one fix-subagent per issue in parallel (partition by file), then reports what was fixed/how and ignored/why. Diff-scoped by default, like CodeRabbit. Ships reviewer prompts + CLI recipes + report template.
multi-agent-codebase-audit	Audit a whole codebase with a fleet of agents: map the repo, partition it into review slices, fan out one+ auditor subagent per slice (multi-lens on high-risk slices), then reduce with a cross-cutting/architecture critic + a completeness critic (loop-until-dry), before triage → parallel fix → report with a coverage ledger (so "audited everything" is honest). Whole-repo sibling of dual-agent-review (reuses its triage/fix/report); scales agents to repo size. Ships partition heuristics + auditor/critic prompts + ledger template.
meaningful-linear-issues	Create complete, analytics-ready Linear issues in one pass instead of metadata-bare ones. Never guess the project (resolve from context → confirm via list_projects → ask); treat the "optional" fields as opt-out — proactively fill priority, estimate, labels (a Type + an Area from the existing taxonomy, not invented), milestone, assignee, cycle, and relations, discovering valid values via the Linear list_* tools. Covers the estimate-snaps-to-scale + cycle-auto-joins gotchas, a description template, and a pre-flight checklist before save_issue.

Repository layout

skills/
  <skill-name>/
    SKILL.md         # entry: frontmatter (name + description) + the skill body
    reference/*.md   # optional — deep docs the skill links to, loaded on demand
    scripts/*        # optional — runnable helpers the skill invokes
    templates/*      # optional — boilerplate the skill copies
README.md
AGENTS.md            # conventions for authoring skills in this repo

The skills CLI auto-discovers skills/<name>/SKILL.md (flat) and skills/<category>/<name>/SKILL.md (categorised) — no manifest required.

A skill is a directory, not just one file: SKILL.md is the entry point, and only its body loads when the skill fires. Anything else (reference docs, scripts, templates) loads/runs only when SKILL.md references it — keep SKILL.md focused and push heavy material into siblings (progressive disclosure). The whole directory installs together, so relative links keep working.

Adding a skill

npx skills init skills/<skill-name>   # scaffolds skills/<skill-name>/SKILL.md

See AGENTS.md for the conventions (what makes a good description, how to keep the body focused, etc.).

License

MIT © Stealth Engine

---

_{Keywords: agent skills · Claude Code skills · Claude skills · AI agent context · skills.sh · Cursor skills · iOS 26 Safari Liquid Glass · prompt engineering.}