sindre-ai

frontend-task

Orchestrates the frontend agent team to plan, review, implement, and review frontend changes

sindre-ai 7 1 Updated 4mo ago
GitHub

Install

npx skillscat add sindre-ai/maskin/frontend-task

Install via the SkillsCat registry.

SKILL.md

Frontend Task Orchestration

You are orchestrating a frontend task using a team of specialized agents. Follow this workflow exactly.

Workflow

1. Assess the Task

Read the user's request and classify it:

  • Trivial: typo fix, single-line change, obvious one-file fix → skip to Phase 2
  • Non-trivial: new feature, multi-file change, UI addition, behavior change → start at Phase 1

2. Create the Team

Create a team with these three agents (defined in .Codex/agents/):

  • frontend-planner — explores codebase and creates implementation plan
  • frontend-reviewer — reviews plans and implementations for quality
  • frontend-implementer — writes the code following approved plans

3. Phase 1: Plan + Review (skip for trivial tasks)

Create these tasks:

Task A: "Plan: {description of the frontend task}"

  • Assign to: frontend-planner
  • The planner will explore existing components, hooks, and patterns, then produce a structured plan

Task B: "Review the plan" (depends on Task A)

  • Assign to: frontend-reviewer
  • The reviewer checks the plan for component reuse, placement consistency, and rule compliance
  • If the reviewer sends feedback, the planner addresses it and the reviewer re-checks
  • Max 2 feedback rounds — after that, the reviewer approves with notes and we proceed

4. Phase 2: Implement + Review

Task C: "Implement the approved plan" (depends on Task B, or no dependency for trivial tasks)

  • Assign to: frontend-implementer
  • The implementer follows the approved plan exactly, then runs lint + type-check

Task D: "Review the implementation" (depends on Task C)

  • Assign to: frontend-reviewer
  • The reviewer checks the code against the rules and the approved plan
  • If the reviewer sends feedback, the implementer fixes and the reviewer re-checks
  • Max 2 feedback rounds — after that, the reviewer approves with notes

5. Finalize

When all tasks are complete:

  1. Run pnpm lint and pnpm type-check as a final safety net
  2. Summarize what was done for the user
  3. List any "approved with notes" items that should be addressed later

Important Notes

  • The agents communicate directly with each other for feedback loops — you don't need to relay messages
  • Each agent reads .Codex/rules/frontend.md and apps/web/AGENTS.md at startup
  • The reviewer is the quality gate — nothing ships without reviewer approval
  • If the task involves adding a new shadcn/ui component, use: cd apps/web && npx shadcn@latest add <component> --yes
  • For tasks that touch both frontend and backend, only handle the frontend portion here