review
Web Design Guidelines Review (Subagent)
Step 1: Determine Target
With argument: Use it directly.
Without argument: Determine from session context:
- Clear single target (one commit, one file) → use it
- Ambiguous (multiple commits, unclear scope) → ask user via
AskUserQuestion
Ask when unclear. Use AskUserQuestion with options:
Always include:
- All session changes — all commits made in this conversation (if multiple)
- Whole feature branch — diff of current branch vs main
- Last commit (hash) — most recent commit only
- Uncommitted changes — only if they exist
Add context-specific options if identifiable:
- Specific component/file discussed in session
- Particular commit mentioned earlier
- Directory being worked on
- Spec/design file referenced
Step 2: Find Design System Path
Quick glob (don't read):
glob: "docs/**/design-system.md"
glob: "**/DESIGN_SYSTEM.md"
Step 3: Spawn Agent
Do NOT gather content yourself. Just pass identifiers — the agent gets the content.
tool: Agent
parameters:
description: "UI review: [target]"
subagent_type: "web-design-guidelines:ui-reviewer"
prompt: |
Review: [TARGET IDENTIFIER - e.g. "commit abc123", "src/Button.tsx", "uncommitted changes"]
Design system: [PATH or "none"]
Context: [BRIEF SESSION CONTEXT - what feature, what spec, focus areas]
Don't:
- Run
git difforgit show - Read file contents
- Write diffs to temp files
Do:
- Pass commit hash, file path, or "uncommitted changes"
- Pass design system path
- Pass brief context (feature name, spec reference, focus areas)
Step 4: Return findings to user
More from mrclrchtr/skills
agent-orchestrator
Orchestrate complex work via a phase-gated multi-agent loop (audit → design → implement → review → validate → deliver). Use when you need to split work into subsystems, run independent audits, reconcile findings into a confirmed issue list, delegate fixes in clusters, enforce PASS/FAIL review gates, and drive an end-to-end validated delivery. Do not use for small, single-file tasks.
38git-commit
Creates a commit: detects conventions, stages intentionally, writes a clear subject, add a concise body when useful, and commits.
22agent-orchestrator-standalone
Orchestrate complex work via a phase-gated multi-agent loop (audit → design → implement → review → validate → deliver). Use when you need to split work into subsystems, run independent audits, reconcile findings into a confirmed issue list, delegate fixes in clusters, enforce PASS/FAIL review gates, and drive an end-to-end validated delivery. Do not use for small, single-file tasks.
16commit
Creates a commit with repo-matching style and intentional staging.
5web-fetch-to-markdown
Fetches and normalizes http(s) web pages into clean Markdown for LLM ingestion. Use when a task includes a URL, needs to fetch docs or asks to convert web docs/articles/pages into Markdown for summarizing, quoting, diffing, or saving.
5web-design-guidelines-design
Use when creating, redesigning, or restyling a UI and Codex should establish a clear design direction before implementation.
5