[nevridge]

Summary

Adds a /ux-planner skill for answering UX and frontend design questions before building, adapted from the SalesPower skill of the same name.

The skill runs codebase-researcher on the relevant frontend area, passes the findings to ux-advisor, and returns one of three outcomes:

• No UI work: the question doesn't need UI changes.
• Design conflict: the user explicitly asked for something the advisor recommends against. The skill quotes the conflict and offers three options with AskUserQuestion.
• Plan: a structured implementation plan built from the advisor's output sections, plus an offer to post decisions to the source issue.

It accepts a question or an issue reference (/ux-planner #307).

Changes

• .claude/skills/ux-planner/SKILL.md (new). Adapted to TaskFlow's advisor sections (STYLE_ADDITIONS, TESTING_CRITERIA), conventions and design philosophy. Reviewed against Anthropic's skill authoring guidance:
  • The description leads with when to use the skill.
  • $ARGUMENTS is used in the body.
  • Repeats of CLAUDE.md content were removed.
  • Only specific trigger phrases are kept.
  • allowed-tools: Bash(gh issue view *) is pre-approved.
• .claude/agents/ux-advisor.md:
  • Notes that it can be invoked from /ux-planner with only a question and research findings.
  • Restricts DESIGN_CONFLICTS to directions the request explicitly asks for. Notes, exceptions and resolved choices go in other sections. Before this change, the advisor sometimes filled the section with non-conflicts, which made the skill stop for user input unnecessarily.

Testing

No app code changed, so there are no unit tests. claude plugin validate .claude/skills passes. Ran the skill in fresh headless sessions (claude -p):

Prompt	Expected	Result
/ux-planner #307	Full plan	Pass (after the DESIGN_CONFLICTS fix; the first run stopped on non-conflicts)
Backend-only question (index on TaskItems.JournalDate)	No UI work	Pass
Confirm dialog before completing a task	Conflict discussion	Pass

Not tested:

• Automatic triggering from plain-language prompts.
• Running the advisor on Haiku.

The #307 run also surfaced #308 (journal updates may clear unsent task fields).

🤖 Generated with Claude Code