Agent skills

Skills are local instruction bundles for a specific task. Alfred can pass them to Claude Code, Codex, or OpenCode through its role-scoped prompt injector. The default setup installs only the first-party skills that improved deterministic task results. Other first-party and third-party packs remain explicit installs.

Full guide at docs/SKILLS.md. Highlights:

Where the default starter set lives

After Alfred Desktop’s full local setup, or after alfred skills install --starter in the CLI-only path, the first-party starter skills land here. Optional vendored and fetched packs install beside these only when you run their explicit install commands.

~/.claude/skills/
├── spec-to-issues/SKILL.md
├── review-security/SKILL.md
└── add-observability/SKILL.md

First-party starter set

These are installed by Alfred Desktop during full local setup and by alfred skills install --starter in the CLI-only path. Each passed two held-out paired tasks with no skill-arm failure, regression increase, or review-finding increase. The install is local, offline, and deterministic.

Skill	Source	Used by	Why
spec-to-issues	Alfred first-party	planner	Converts specs into issue queues
review-security	Alfred first-party	reviewer, feature-dev	Review checklist for risky code
add-observability	Alfred first-party	feature-dev, ops-watch	Logging and metrics prompts

write-tests, migrate-dependency, and changelog-and-release-notes remain available as explicit first-party installs. write-tests was neutral. The migrate-dependency and changelog-and-release-notes skills each improved one task but did not meet the two-task evidence floor. Alfred does not install or inject them by default.

Curated optional packs

These are part of the registry, but not part of --starter. Install the vendored packs explicitly when you want the heavier specialist lenses. They are still local offline copies once installed. Fetched packs require --yes because they reach out to third-party sources at install time.

Skill	Source	Used by	Why
code-review-and-quality	addyosmani/agent-skills	feature-dev, reviewer, fixer	Multi-axis review
security-and-hardening	addyosmani/agent-skills	feature-dev, reviewer	Security-specific lens
frontend-ui-engineering	addyosmani/agent-skills	feature-dev	Production UI patterns
debugging-and-error-recovery	addyosmani/agent-skills	bug-triage, ops-watch	Systematic root-cause path
vercel-react-best-practices	vercel-labs/agent-skills	feature-dev	React and Next.js performance guardrails
gstack	garrytan/gstack	optional: reviewer, triage, e2e-runner	CLI-first review, QA, and ship flow
headroom	headroomlabs-ai/headroom	optional	Token and context inspection

Install

alfred skills list
alfred skills install --starter

alfred skills install code-review-and-quality
alfred skills install security-and-hardening
alfred skills install frontend-ui-engineering
alfred skills install debugging-and-error-recovery
alfred skills install vercel-react-best-practices

# Optional fetched packs. These require explicit confirmation because they
# pull from third-party sources at install time.
alfred skills install gstack --yes
alfred skills install headroom --yes

For the full CLI reference, see docs/SKILLS.md#the-alfred-skills-command.

Security note

Skills run with the same permissions as the coding CLI. They can read files the local user can read, write files the user can write, run shell commands, invoke tools, and use the network. Treat a new skill like any other executable dependency:

1. Read the SKILL.md.
2. Skim the scripts the skill might invoke.
3. Run a Snyk / CodeQL scan on unfamiliar sources.
4. Pin to a specific commit when installing from a third-party tap.

A worktree separates git changes. It does not isolate the process from the home directory, credentials, or network. Use a dedicated operating-system user, virtual machine, or container, plus egress controls where required.

Anti-recommendations

• Anything that auto-publishes (auto-tweet, auto-deploy, auto-merge). Use as draft-then-review only.
• Skills that fork to the network without explicit allowlists. Network egress from a worktree is a known agent attack vector.
• Skills you have not read. Skills are markdown. Read them.

Where skills live in the framework’s mental model

Skills are local operator assets, not a hosted service. Alfred ships a curated first-party starter set for the default engineering fleet, plus optional vendored and fetched packs for teams that want heavier tools. The registry records source, license, install method, and default roles so a fleet can stay batteries-included without hiding third-party code from the operator.