skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
← back to all skills
arla6ka/design.how272 installs

design-system-boss

Use for vague whole-app design system asks on an app that already ships UI, like "our UI is a mess, fix it", "we need a design system", "we hardcode colors everywhere", "make the buttons consistent for an upstream PR", "document our whole design system" or "check the app before I ship". Triages the app, then runs the design.how skills in order. One named screen goes to ui-review, one component to component-docs, an empty repo to build-design-system.

How do I install this agent skill?

npx skills add https://github.com/arla6ka/design.how --skill design-system-boss
view source ↗

Is this agent skill safe to install?

  • Gen Agent Trust Hubpass

    The skill acts as a design system coordinator, performing repository analysis and delegating tasks to other specialized design skills. It uses a shell script for repository triage and can execute standard CLI tools. It possesses an attack surface for indirect prompt injection via repository-level configuration files and source code processing.

  • Socketpass

    No alerts

  • Snykpass

    Risk: LOW · No issues

What does this agent skill do?

Design system boss

  1. Resume from .design-system/boss/state.md if it exists (references/state.md, Resuming). Else run scripts/triage.sh (references/triage.md).
  2. Pick the state, the intent and the foundation from the signals (references/triage.md). With small_app yes, take the fast path (references/routes.md, Small app).
  3. Pick the route from the table below. It ends check-first unless the ask names the migration (references/routes.md, the rules at the top).
  4. Pick the run branch, set the phase caps, probe the host, write the state file (references/coordinator-path.md, steps 6 to 9).
  5. Send the Frame with every default applied. An unanswered question keeps its default as default (unanswered), and the run goes on (references/triage.md, The one question).
  6. Run the route's steps through subagents, one writing step at a time (references/delegation.md). Decided defaults land before any docs are written.
  7. Check each return on its files and record a verdict (references/delegation.md, Checking a return).
  8. Close: triage again, the clean-clone check, captures, then close.md and one found-not-fixed table (references/state.md).
  9. Send the Report section as the final message, unanswered defaults first and the migration offered with its size.

Never ship:

  • Product code written by the boss on a host that can spawn subagents.
  • A migration past the safe moves without clearance, or a "fix it" read as clearance.
  • A docs or spec commit made before the decided defaults landed.
  • A new visual direction nobody chose through the design-source question.
  • A number in the report that no row in close.md holds.
  • A handback with a worker still running.
  • A run that waited or stopped on a question it had a default for.

When the boss runs as a subagent, it spawns workers as foreground calls, several in one message to run them side by side, and never ends its turn with a background worker live. A worker left running keeps writing into a repo nobody checks.

The five siblings do the work: build-design-system, migrate-design-system, token-mapping, ui-review and component-docs, each called through its "When a coordinator calls it" section. Sibling paths start at <skills>. Terms such as surface, gate, clearance and the status words mean what build-design-system/references/run-record.md (Terms) says, and its standing orders bind the boss too. Read references/coordinator-path.md first. It lists each step and the file to open.

Output

Done means the route's last step has returned, every step has a status and verdict in .design-system/boss/state.md, and the report is written from files. Checkable:

  • triage/signals.tsv and triage/after/signals.tsv came from scripts/triage.sh, and the route is the routing table's, or a decision row says why it differs.
  • Every step has a verdict naming the file that proves it. Every number in the report is a row in close.md with its unit and source.
  • Every gate from every sibling record is in the state file's Gates table (references/state.md, The state file), and every unanswered one reads default (unanswered).
  • On a writing route, the check exits 0 on a clean clone of the run branch, or the repo's own lint, typecheck and build do on a minimal footprint. Every changed surface has captures and a montage row, every identical-value swap has its proof, and every pilot trap is fixed with numbers or gated with its measurement. git status before and after is saved, .design-system/tmp/ is gone, and every untracked path is committed or named in a decision row.
  • The report ends with one Found, not fixed table, ranked, at most 30 rows.

Your final chat message is the state file's Report section, verbatim (references/state.md, The handoff report). If a stop condition below holds, return the condition, what finished, the state file path, and the smallest reply that unblocks it, such as "Say which app: apps/web or apps/admin."

Inputs

InputIf missing
The askRequired. "Our UI is a mess" is enough
The repoStop without one. Read-only access allows only Audit and Review
Target app, in a monorepoThe one question, default the app with the most routes. Triage every app meanwhile
IntentFrom the ask's words, per references/triage.md
BudgetThe Frame's Go, <budget> reply, else the host's session, else 2 hours, split per references/routes.md (Budget)
The sibling skillsIn <skills>. A missing one stops only its steps, and the report names it
Subagents, a browser, a shellProbe per references/delegation.md
AGENTS.md or CLAUDE.md, the person's words, saved memoryThey win over every skill default, branch and push policy included. "Work on my current branch" makes it the run branch. Pass them into every brief

Clear the migration

Clearance comes from an ask that names the migration or a reply to the handoff's offer (build-design-system/references/run-record.md, Terms). A fallback intent and a complaint such as "fix it" never count. With clearance, migrate edits only after its audit plan.md is re-pinned and reconciled with the build's registry.json, and migrate's rolling window runs inside the machine budget on disjoint paths. Without it, the route ends check-first. An ask mid-run follows references/routes.md (the rules at the top).

Routing table

State, weak and intent come from references/triage.md. references/routes.md has each route's steps. Copy them word for word, and keep a dropped step listed as skipped (<reason>).

StateIntentRoute
installedreviewReview, with the installed system's SKILL.md as the criteria. Wins over every row below
installedauditAudit, with the installed system as the target
installedany otherInstalled system
system reporeview, audit, componentThat intent's row below
system repoany otherSystem repo
emptyanySeed
none, or drifting and not weakbuildBuild
drifting, weakbuildHarden, stated in the Frame
any, with a PR or upstream in the asknamed familiesNamed families, minimal footprint. Wins over every row below
drifting or settled, weakfull, harden, adoptHarden, then Full from step 5. A values intent runs Values first. Wins over later rows
none, drifting or settled, not weakhardenHarden. With no harden_dirs, Build
none or drifting, not weakfullFull
settled or documented, not weakfull, adoptAdopt
drifting, not weakvaluesValues
settleddocsDocument
anycomponentComponent
anyreviewReview
anyauditAudit

An installed system is the source of truth, so no route seeds, builds or hardens over it. It grows only by a gap recorded for its owner (references/triage.md, The installed system). When the ask mentions a PR or upstream, or the repo looks like someone else's, the route runs as <route>, minimal footprint (Footprint in run-record Terms). When the ask and the state disagree, the state wins and the Frame says so. "Migrate us" on state none is Build.

Boundaries

May decide the route, run branch, phase split, which read-only steps run side by side, one retry of a failed step, defaults for any sibling input the ask left out, and skipping a step whose output is current, citing its record.

Ask, with the default applied, and go on:

  • A dirty checkout or someone else's branch on a writing route. Default: the run branch in its own worktree from HEAD, so the checkout stays as found. Never stash.
  • Read-only access on a route that writes. Default: Audit.

Stop only when there is no repo, or a sibling returns its stop shape and the next step needs its output.

Never loosen a sibling's predicate or edit a sibling skill. "Make it modern" with no design source becomes a gate with the default "keep the current look" (references/triage.md, Standing questions).

Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.

<a href="https://skillzs.dev/skills/arla6ka/design.how/design-system-boss">View design-system-boss on skillZs</a>