SDLC Agent-Skill Wiki

addy-agent-skills

7 min read

Addy Osmani — Agent Skills

Workflow — the primary skill per SDLC stage this framework runs, top to bottom (folded and off-stage steps omitted). Validate and Review are sibling gates that both run after Implement.

flowchart TB
    A["addy-interview-me<br/><i>Align</i>"] --> S["addy-spec-driven-development<br/><i>Specify</i>"] --> P["addy-planning<br/><i>Plan</i>"] --> I["addy-incremental-implementation<br/><i>Implement</i>"]
    I --> V["addy-browser-testing<br/><i>Validate</i>"]
    I --> R["addy-code-review<br/><i>Review</i>"]
    V --> Rel["addy-shipping<br/><i>Release</i>"]
    R --> Rel

Production-grade engineering skills for AI coding agents by Addy Osmani — a pack of 24 skills, 4 review personas, and 8 slash commands that encode the workflows and quality gates senior engineers use, organized around the whole product lifecycle. MIT-licensed; distributed as a Claude Code plugin and portable across Cursor, Gemini CLI, Antigravity, OpenCode, Windsurf, Copilot, Kiro, and Codex.

  • Install (Claude Code): /plugin marketplace add addyosmani/agent-skills then /plugin install agent-skills@addy-agent-skills.
  • Design thesis: agents default to the shortest path (skipping specs, tests, security) — skills impose senior-engineering discipline. Bakes in Software Engineering at Google concepts: Hyrum’s Law, the Beyoncé Rule, the test pyramid, change sizing, review-speed norms, Chesterton’s Fence, trunk-based development, Shift Left, code-as-liability.

Where gsd is a single end-to-end workflow engine and matt-pocock-skills is a loose personal toolkit, Agent Skills is a lifecycle-complete, phase-structured pack: slash commands map 1:1 to six phases, a meta-skill (addy-using-agent-skills) routes work to the right skill, and every skill carries an anti-rationalization table and Red Flags (pattern-anti-rationalization). It is the broadest framework in this wiki — the first to give the stage-release stage substantial multi-capability evidence.

Distinctive mechanisms

  • Anti-rationalization + Red Flags in every skill — pattern-anti-rationalization tables of excuses agents use to skip steps, each rebutted, plus warning signs.
  • Review personas with parallel fan-out — addy-shipping (/ship) fans out addy-code-reviewer, addy-security-auditor, addy-test-engineer concurrently, then merges a go/no-go (pattern-parallel-persona-review). Personas never invoke personas.
  • Verification is non-negotiable — every skill ends with evidence requirements (tests passing, build output, runtime data); “seems right” is never sufficient.
  • Progressive disclosure — SKILL.md is the entry point; reference checklists load only when needed.
  • Human checkpoint at each phase — commands stop between phases (except /build auto, which runs the whole plan in one approved pass, still test-driving and committing each task individually).

Lifecycle: phase ↔ command ↔ skill ↔ canonical stage

PhaseCommandPrimary skill(s)Canonical stage
Define/specaddy-interview-me, addy-idea-refine, addy-spec-driven-developmentstage-align → stage-specify
Plan/planaddy-planningstage-plan
Build/build (/build auto)addy-incremental-implementation + addy-tddstage-implement
Verify/testaddy-tdd, addy-browser-testing, addy-debuggingstage-implement / stage-validate
Review/review, /code-simplify, /webperfaddy-code-review, addy-code-simplification, addy-security, addy-performancestage-review
Ship/shipaddy-shipping + fan-out personasstage-release

Note on Verify vs Review. Addy was the first framework to cleanly separate does it work (Verify — testing, debugging) from is it good (Review — code review, security, performance, simplification). This partition — confirmed by gstack as a second framework (2026-07-05) — promoted stage-review out of stage-validate: Addy’s Review phase (plus its /ship persona fan-out) implements stage-review; its Verify capability addy-browser-testing stays in stage-validate.

Capabilities

Skills — Define (align → specify)

Skills — Plan

  • addy-planning — decompose specs into small, verifiable, dependency-ordered tasks.

Skills — Build

Skills — Verify

Skills — Review

Skills — Ship

  • addy-git-workflow — trunk-based development; atomic commits; semver; changelogs.
  • addy-ci-cd — Shift Left; feature flags; quality-gate pipelines.
  • addy-deprecation — code-as-liability; compulsory vs advisory deprecation; zombie-code removal.
  • addy-documentation — ADRs, API docs, inline docs — document the why.
  • addy-observability — structured logging; RED metrics; OpenTelemetry; symptom-based alerting.
  • addy-shipping — pre-launch checklists; staged rollouts; rollback; the /ship fan-out.

Skills — Meta

Agent personas (sub-agents)

Commands (thin phase entrypoints, catalogued here)

Each command invokes the skill(s) above and is documented in the phase table; they carry no independent workflow beyond skill activation, so they are not given separate capability pages.

  • /spec, /plan, /build (+ auto), /test, /review, /code-simplify, /webperf, /ship.

Reference checklists (bundled, pulled in on demand)

  • definition-of-done.md, testing-patterns.md, security-checklist.md, performance-checklist.md, accessibility-checklist.md, observability-checklist.md, orchestration-patterns.md.

Artifacts produced

Patterns applied

See Also

Graph View

Backlinks