RepoDaily · 2026-07-07 · Security tool

Agent Skills: Engineering Guardrails That Make AI Coding Agents Ship Safer Code

#4 Security tool Shell +1,114 addyosmani/agent-skills Open repository

Addy Osmani's pack of 24 Markdown skills enforces spec-driven development, test-driven workflows, five-axis code review, and security hardening across 70+ AI coding agents — including autonomous /build auto mode.

Repo typeSecurity tool
Best forEngineering teams whose AI coding agents (Claude Code, Cursor, Codex, Copilot, Cline) skip tests, security checks, or code reviews unless explicitly told not to
Risk levelLow — skills are Markdown instruction files loaded into agent context; no background daemons or network calls
Time to evaluate20–30 minutes to install via npx skills CLI and run a /spec → /build → /test cycle

Primary question: Are your AI-generated commits passing tests, security scans, and code review — or are you hoping they are?

93/100

RepoDaily adoption score

RepoDaily rates this as 93/100 (strong) for adoption: evidence, installation path, production risk, differentiation, license clarity, and AI/agent fit are scored from the article sources and adoption notes.

Directional score from RepoDaily sources and adoption notes, not a benchmark.Risk: Low
96Evidence quality

4 source(s) across 3 source category/categories, plus a RepoDaily-specific evidence module when available.

100Installability

6 workflow step(s), 5 next-action step(s), and 3 command/install signal(s) were detected.

78Maintenance confidence

Trending momentum is +1,114 stars, with maintenance/release/issue signals counted when present.

100Production readiness

Risk is marked low, with 6 security note(s) and 4 explicit skip condition(s).

100Differentiation

3 opportunity lens item(s), 4 alternative(s), and 3 type-specific section(s) support differentiation.

82License clarity

License source or license wording is present.

84Agent / AI fit

6 AI/agent-related signal(s) were detected in the article text and metadata.

Project overview

Agent Skills is a curated pack of 24 production-grade engineering workflows packaged as Markdown files that AI coding agents follow as step-by-step processes rather than reference documentation. Created by Addy Osmani, the project maps senior-engineer practices — spec-driven development, test-driven development, incremental implementation, code review, security hardening, and shipping — onto eight slash commands that cover the full development lifecycle from idea to production deploy.

The core insight is structural: each skill is a SKILL.md file with a defined anatomy — Overview, When to Use, Core Process, Common Rationalizations, Red Flags, and Verification exit criteria. When loaded into an agent's context, the skill dictates what the agent must do and what evidence it must produce before declaring a task complete. The security-and-hardening skill and the security-auditor agent persona are specifically designed to catch vulnerability patterns before merge.

Installation targets 70+ agents through the open skills CLI from vercel-labs/skills, with native integrations for Claude Code (marketplace plugin), Cursor (.cursor/rules/), and any agent that accepts Markdown system prompts. The /build auto command goes further: it generates a plan from an approved spec, implements every task in a single autonomous pass, commits each task individually, and pauses on failures or risky steps — preserving verification while removing the human bottleneck between tasks.

Problem it solves

  • AI coding agents frequently skip writing tests, producing code that appears functional but lacks verification — the test-driven-development skill enforces red-green-refactor with exit criteria
  • Security checks are routinely omitted because agents optimize for completing the task, not hardening it — the security-and-hardening skill and security-auditor.md agent persona address this directly
  • Code review is treated as optional by most agents — the code-review-and-quality skill runs a five-axis review (correctness, security, performance, maintainability, test coverage) before merge
  • Large changesets generated in a single pass are hard to review and rollback — incremental-implementation forces one-slice-at-a-time commits with individual test runs
  • Skills shipped as reference docs get ignored — agent-skills frames each skill as an actionable process with verification, not background reading

How it works

  1. Install the skills pack: run npx skills add addyosmani/agent-skills to load all 24 skills, or use npx skills add addyosmani/agent-skills --list to browse first
  2. Select individual skills if needed: npx skills add addyosmani/agent-skills --skill code-review-and-quality installs just the five-axis review skill
  3. For Claude Code, use the marketplace: /plugin marketplace add addyosmani/agent-skills then /plugin install agent-skills@addy-agent-skills — or clone locally and run claude --plugin-dir /path/to/agent-skills
  4. For Cursor, copy any SKILL.md file into .cursor/rules/ — the agent picks it up as a project rule
  5. Load the using-agent-skills meta-skill first — it contains a flowchart that maps task types to the appropriate skill, preventing context waste from loading everything at once
  6. Use slash commands to trigger lifecycle phases: /spec to define requirements, /plan to break work into atomic tasks, /build to implement incrementally, /test to prove correctness, /review to run five-axis review, /ship to deploy

Slash Commands and Skill Activation Map

The project exposes 8 slash commands that map directly to lifecycle phases. Each command activates the relevant skills automatically — designing an API triggers api-and-interface-design, building UI triggers frontend-ui-engineering. The /build auto variant generates a plan and implements every task in a single approved pass, committing each task individually and pausing on failures. The session-start hook (hooks/session-start.sh) injects the using-agent-skills meta-skill into every new Claude Code session, with a regression test at hooks/session-start-test.sh that validates the JSON payload both with and without jq installed.

Fastest Evaluation Path

  • Run npx skills add addyosmani/agent-skills --list to browse all 24 skills before committing to an install
  • Start with the minimal trio: spec-driven-development, test-driven-development, and code-review-and-quality — these cover the most critical quality gaps
  • Test /build auto on a small feature: approve the generated plan once, then observe whether per-task commits, test runs, and pause-on-failure behavior work as documented
  • Run bash hooks/session-start-test.sh before opening any PR that touches the hook — expected output is 'session-start JSON payload OK', non-zero exit on failure

What to Verify Before Adopting

  • Confirm your agent supports Markdown rules files or system prompts — all 70+ agents in the skills CLI registry do
  • Check that the skill anatomy (Overview, When to Use, Process, Common Rationalizations, Red Flags, Verification) matches your team's review expectations
  • Review the MIT license — no commercial restrictions, but no warranty either
  • Test context consumption: the docs recommend loading only skills relevant to the current task, not all 24 at once, to avoid wasting context window

Who should pay attention?

Good fit if

  • Teams using Claude Code, Cursor, Codex, or Copilot who want consistent quality gates without writing custom rules from scratch
  • Projects where AI-generated code needs enforced test-driven development before merge
  • Engineering leads who want a five-axis code review (correctness, security, performance, maintainability, test coverage) applied to every AI-authored change
  • Developers evaluating /build auto for autonomous implementation with safety pauses on risky steps
  • Anyone who has experienced an AI agent rationalizing why it skipped a step — the Common Rationalizations section directly addresses this pattern

Skip for now if

  • Teams not using AI coding agents — the skills have no effect without an agent to load them
  • Projects that already have mature, enforced CI/CD pipelines with automated security scanning and may find overlapping coverage
  • Developers who prefer minimal system prompts and find 24 skills too heavy for their context window — selective loading mitigates but does not eliminate this
  • Organizations requiring SOC 2 or compliance audited tooling — this is a community project under MIT license with no formal security certification

Risks and cautions

Low

Skills are static Markdown files with no runtime code execution, no network calls, and no background processes. The session-start hook is a shell script with a regression test. MIT license carries no commercial restrictions.

  • No compiled binaries, no npm dependencies at runtime — skills are text files loaded into agent context
  • The hooks/session-start.sh hook is testable via hooks/session-start-test.sh with documented expected output
  • MIT license permits commercial use, modification, and distribution with no warranty
  • Context window consumption is the main operational cost — mitigated by context-aware loading guidance in the docs
  • The security-and-hardening skill is loaded before merge in the recommended full-lifecycle setup, running alongside code-review-and-quality
  • A dedicated security-auditor.md agent persona in the agents/ directory provides specialized vulnerability detection when invoked
  • The code-review-and-quality skill performs five-axis review that includes a security axis — not just style or correctness
  • The /build auto command pauses on risky steps rather than proceeding autonomously through potentially dangerous operations
  • Skills include Red Flags sections that define signs the skill is being violated — providing auditable criteria for whether security practices were followed
  • The session-start hook regression test (hooks/session-start-test.sh) validates the JSON payload integrity, preventing silent corruption of the meta-skill injection

Alternatives to compare

ApproachWhen to useTrade-off
vercel-labs/skills (Skills CLI)
You want the installer tool itself rather than a specific skill pack — the CLI distributes many skill collectionsFree, MIT-style open source
Claude Code built-in plugin marketplace
You only use Claude Code and want official plugins without external dependenciesIncluded with Claude Code subscription
Cursor .cursorrules files
You want lightweight per-project rules without a full lifecycle skill frameworkFree, included with Cursor
Custom system prompts
You have unique engineering standards that no off-the-shelf pack coversFree, requires manual authoring and maintenance

What this trend reveals

Extend with domain-specific security skills

The CONTRIBUTING.md defines a clear anatomy and quality bar for new skills. A team could author a skill for OWASP Top 10 scanning, dependency vulnerability triage, or secrets detection that plugs into the existing /review and /ship commands.

Follow docs/skill-anatomy.md, create a kebab-case directory under skills/, add SKILL.md with YAML frontmatter, and justify the gap against existing skills in the PR description.

Use the meta-skill flowchart for agent onboarding

The using-agent-skills meta-skill maps task types to appropriate skills via a flowchart. This can serve as a diagnostic tool when evaluating whether an AI coding agent is applying the right practices to a given task type.

Load using-agent-skills into the agent context, then ask it to identify which skill applies to a sample task — verify the mapping matches the documented flowchart.

Adopt /build auto for controlled autonomous sprints

Teams hesitant about fully autonomous AI coding can use /build auto as a graduated step: approve the plan once, then let the agent implement with per-task commits, test verification, and pause-on-failure safety.

Run /build auto on a low-risk feature branch, observe the pause behavior on a deliberately failing test, and confirm each task produces an individual commit.

Best next action

Install the minimal trio and run a /spec → /build → /review cycle

Start with three skills that address the most critical quality gaps: spec-driven-development, test-driven-development, and code-review-and-quality. This gives you immediate enforcement of requirements definition, test verification, and five-axis review without loading all 24 skills.

  1. Run npx skills add addyosmani/agent-skills --skill spec-driven-development --skill test-driven-development --skill code-review-and-quality
  2. Open your AI coding agent and run /spec on a small feature request
  3. Once the spec is generated, run /build to implement incrementally with test verification
  4. Run /review to trigger the five-axis code review before considering the change mergeable
  5. Check the Verification section output — if any exit criteria are unmet, the skill was not correctly applied

RepoDaily verdict

Agent Skills is the most structured attempt yet to make AI coding agents follow senior-engineer practices consistently. The 24-skill pack, slash-command lifecycle, and /build auto mode address the real failure mode of AI-assisted development: agents that generate plausible code while skipping tests, security checks, and review. The Markdown-only architecture keeps adoption risk near zero, and the skill anatomy with Common Rationalizations and Red Flags is a pragmatic pattern that other rules frameworks should adopt.

Sources