Aeo Claude

Use when a multi-step task in Claude Code should be planned by a coordinator and implemented by parallel subagents: multi-agent planning for large refactors, multi-component features, or any work worth delegating to independent Opus subagents with a tracked task list. The coordinator enters plan mode, gets the breakdown approved, delegates each task, and verifies every result; subagents work independently and report back. When teammates must message each other mid-task, use ultraplan-teams instead.

Use when validating a plan, code change, configuration, or any other work product before proceeding — checking the preceding context for errors, unvalidated assumptions, alignment issues, gaps, and improvement opportunities, and producing a machine-parseable summary of what it found. Trigger phrases include "ultrareview", "review this plan", "check my assumptions", "validate this before we build", and "what did we miss". Reports findings only; the companion `ultrareview-fix` skill applies them.

Use when moving, importing, copying, or migrating a Claude Cowork session, conversation, or project from one Windows machine to another, or when troubleshooting symptoms of a broken import like "session shows blank", "only the latest messages show", "scratchpad files don't open", "can't scroll past the last compaction", or "Loaded N messages (truncated via tail/compaction)" in the Cowork log. Delivers full history, working file links, and no truncated-transcript rendering bug, handling the undocumented two-layer compact_boundary truncation filter in app.asar that silently clips imported transcripts. Covers orphan sessions on Windows to Windows under the same Cowork account. Does not handle Cowork Spaces/Projects, Linux/macOS, or cross-account migration.

Use when writing, reviewing, debugging, or rewriting instructions that Claude itself will read — CLAUDE.md files, skill files (SKILL.md, frontmatter descriptions), slash command templates (.claude/commands/*.md), and agent definitions (.claude/agents/*.md) — and when converting docs or runbooks into skills, drafting agent frontmatter from scratch, or asking what is "pulling weight" in a prompt. Symptoms it covers: an agent firing too often or on the wrong things, a prompt that is too long or full of MUST/NEVER/ALWAYS, Claude ignoring sections of a command, agents looping on tools instead of stopping, skills not activating, and descriptions that do not trigger correctly. Not for general code review, MCP server design, hook scripts, or API/SDK questions unless the artifact under discussion is a prompt Claude reads.

Use when a multi-component project in Claude Code needs an agent team whose teammates message each other mid-task over a shared task list, rather than independent subagents: parallel workstreams with file-ownership boundaries, cross-cutting changes where one teammate's output unblocks another, or long builds coordinated with SendMessage. Extends ultraplan with Claude Code's teams feature; the coordinator plans in plan mode, gets approval, creates the team, and verifies. For independent subagents that report back, use ultraplan instead.

Use when working through the findings of a preceding ultrareview validation — systematically addressing the reported errors, alignment issues, gaps, and improvements by priority and applying each fix to the plan or the code. Trigger phrases include "ultrareview-fix", "fix the review findings", "address the ultrareview", and "apply those fixes". Expects an ultrareview summary in the conversation; the `ultrareview` skill produces it.

Use when building autonomous AI agents in Python with the Claude Agent SDK — programmatic agents, tool chains, or multi-agent systems — and when deciding which SDK API or option a task needs. Covers the GTVR execution pattern (gather, take action, verify, repeat), the stateless query() and stateful ClaudeSDKClient APIs, streaming versus batch modes, custom tool creation with the @tool decorator, MCP server integration, hook-based execution control, permission boundaries, API key and OAuth token management, multi-agent coordination, and production deployment strategies. Not designed for basic chat interfaces.

Use when a plan or a code change should go through a code review loop that keeps reviewing and fixing until no actionable findings remain: PR review before merge, plan review before implementation, or validation after a refactor. Runs /ultrareview, applies /ultrareview-fix to the findings, and re-reviews until the review passes or the iteration cap is reached. Unlike a single-pass reviewer, the loop closes its own findings. Usage: /ultrareview-loop [focus_area] [--max-iterations N].

Use when creating a new skill from scratch, modifying or improving an existing skill, packaging one for distribution, or measuring skill performance — running evals to test a skill, benchmarking skill performance with variance analysis, or optimizing a skill's description for better triggering accuracy. Covers SKILL.md frontmatter, progressive disclosure, bundled references, scripts and assets, and the eval, benchmark, and description-optimization pipeline. Trigger phrases include "create a skill", "improve this skill", "run evals on my skill", and "why isn't my skill triggering".

Use when authoring a custom Claude Code slash command, debugging slash command syntax, or converting a repetitive prompt into a reusable shortcut under `.claude/commands/`. Covers the reusable Markdown prompt template format, frontmatter configuration, argument handling, variable substitution, and workflow automation patterns. Trigger phrases include "create a slash command", "make a /command", "custom command", "argument-hint", and "my slash command does nothing".