skills

agent
Guvenlik Denetimi
Uyari
Health Uyari
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 9 GitHub stars
Code Uyari
  • process.env — Environment variable access in skills/mobile-responsive-review/scripts/mobile-scan.mjs
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Agent skills and guardrails for designers who ship production code: Figma fidelity, review gates and the full change workflow.

README.md

Skills

skills.sh

Agent skills for designers who ship production code. They are the guardrails I use to ship with
Claude Code as a founding designer: build a Figma design exactly, keep the design system honest,
and get every change from idea to merged without an agent ever pulling a trigger a human should.

Agents write code quickly. What makes that code mergeable is structure: a fixed order of
operations, a written rulebook, and checks that don't trust the agent or the author. These skills
are that structure, with the company-specific parts removed.

Install

npx skills@latest add Kemdirimakujuobi/skills

Then copy what you need from templates/ into your repo and fill in the
configuration table in AGENTS.md. The skills read your base branch, preflight script, changelog
and docs locations from there, and fall back to sensible defaults when something is missing.

The rules underneath

  • Agents may push branches and open draft PRs. Agents never merge, deploy, or push to the base
    branch.
    A human pulls the trigger. Approval in one message does not carry to the next.
  • One rulebook, three readers. STANDARDS.md is read by the pre-push skill, the review bot and
    the human reviewer. Every rule has an ID, so every finding cites one. No rule ID, no comment.
  • The model finds problems; a deterministic check decides whether they block. Fail closed.
  • The design file is the source of truth. No agent is ever in the token value path.
  • Every mistake becomes a rule. RETRO.md → a skill step → a STANDARDS rule → a CI check.
    starting-a-change reads the latest retro before every branch, so a lesson reaches every
    session after it.

The life of a change

idea
 └─ authoring-a-prd
     └─ starting-a-change
         ├─ figma-import
         ├─ building-components / design-tokens
         └─ figma-build
             └─ pre-push-review
                 ├─ updating-the-changelog
                 ├─ docs-sync-review
                 └─ mobile-responsive-review
                     └─ PR → CI + review bot → a human merges
                         └─ landing-a-design-system-change

Reference

Workflow

  • authoring-a-prd: turns a brief into a PRD with user
    stories, fields, analytics events and numbered acceptance criteria. Every product claim traces
    to a source; anything unconfirmed becomes an open question. Seeding an issue tracker needs
    explicit confirmation. Ships with TEMPLATE.md.
  • starting-a-change: runs before any branch exists.
    Names it <area>/<slug>, cuts it from a fresh base, refuses to stack, and carries the latest
    retro's traps into the plan.
  • pre-push-review: runs before every push. Greps the diff
    against the rulebook, runs two review passes and the local gates, checks size and evidence, then
    calls the three sub-reviews below where they apply.
  • updating-the-changelog: every PR adds a
    customer-facing changelog entry or records Changelog: none — <reason>.
  • docs-sync-review: maps the diff to the help pages it
    makes wrong and drafts the fix in the same PR.
  • mobile-responsive-review: scans added markup
    for what breaks at 360–430px and drafts max-md: fixes that leave desktop unchanged. Ships with
    a scanner script.

Design system

  • building-components: a headless primitive with
    every part is the base, your semantic tokens are the only paint, and the doc contract is the
    finish line.
  • design-tokens: tokens travel from the design file through
    a Git sync and a deterministic CI transform. Agents explain and review token diffs; they never
    produce a value.
  • landing-a-design-system-change: a merge
    is not delivery. Checks the release was declared, the package published, the product bumped,
    and the installed build actually contains the change.

Figma fidelity

  • figma-import: enforces the extraction order. Full node
    first, then spacing, colour, type, radii and every state. The screenshot comes last, only to
    verify against.
  • figma-build: the 1:1 round-trip. Build, render, measure,
    diff, fix, repeat until the measured difference is zero in every state and both themes. Ships
    with MEASURE.md.

Templates

File What it is
AGENTS.md Agent onboarding plus the configuration table the skills read.
WORKING_AGREEMENTS.md The discipline, for people and agents alike.
docs/standards/STANDARDS.md The numbered review rubric, with severity, enforcement tiers, suppressions, waivers and ratchets.
RETRO.md The memory starting-a-change reads before every branch.
docs/decisions/_template.md One decision per file.
.github/workflows/bug-bot.yml + bug-bot-prompt.md A two-job Claude review gate: the model posts findings, a shell step computes the verdict, the rubric is read from the base branch, and it fails closed.

The story of how this setup grew, and what each piece has prevented, is in the release note on my
portfolio: Shipping with guardrails.

License

MIT

Yorumlar (0)

Sonuc bulunamadi