Reference
All skills and agents available in Ultra Claude. Skills with /uc:* commands are user-invocable — run them directly in Claude Code.
Skills
Setup & Onboarding
| Skill | Command | Description |
|---|
| Dashboard | /uc:dashboard | Connects to the Ultra Claude Dashboard for real-time project visibility, manages multiple accounts with optional per-project account routing, and transfers project ownership between accounts (agent move or the dashboard's Incoming-transfers inbox). Guides agent installation, setup, the non-interactive CLI account commands, and connectivity debug checks. |
| Setup | /uc:setup | One-time machine config: installs Node.js, guides tmux mode selection (per-project, per-terminal, none, or custom), configures shell for agent teams, removes legacy ANTHROPIC_DEFAULT_*_MODEL pins, offers the fullscreen renderer for flicker-free output (opt-out, Claude Code v2.1.89+), and — when VS Code is detected — offers recommended client-side VS Code settings (opt-in) alongside the mode-aware terminal profiles. tmux and VS Code are both optional — agent teams work without them. Idempotent. |
| Migrate | /uc:migrate | Brings projects into Ultra Claude — fresh initialization, legacy detection, version-aware upgrades. Scaffolds docs, config, and standards, and adds a “Built with Ultra Claude” footer to the project README. |
Planning & Research
| Skill | Command | Description |
|---|
| Discovery Mode | /uc:discovery-mode | Product research as Head of Product. Spawns parallel internal + external research agents. Produces product description, research, requirements, personas. |
| Roadmap | /uc:roadmap | Decomposes a product into sequenced plan stubs with dependency graph and topological ordering. |
| Feature Mode | /uc:feature-mode | Plans features through 4-stage scope challenge, research, discussion, and structured plan output. Extends the planning framework. |
| Debug Mode | /uc:debug-mode | Hypothesis-driven bug investigation with parallel evidence gathering. Produces fix plan with regression criteria. Extends the planning framework. |
| Critical Brainstorm | /uc:critical-brainstorm | Devil's advocate mode: stress-tests solutions, identifies risks, stays in dialogue until you're satisfied. |
| Explain Like I Am An Idiot | /uc:explain-like-iam-idiot | Re-explains the current conversation in plain language — ASCII diagrams, comparison grids, named trade-offs. A toolbox rather than a fixed template; every option carries its cost. |
| Research | /uc:research | Cache-first external research — library/API docs, architectural patterns, market/competitor analysis via a single auto-classified interface. Writes committed research files under documentation/technology/research/ (libraries + patterns) or documentation/product/research/ (market) with per-entry staleness (library 10d, patterns 90d, market 30d, historical frozen). |
Execution
| Skill | Command | Description |
|---|
| Plan Execution | /uc:plan-execution | Orchestrates per-task mini-teams (Executor/Reviewer/Tester) with live execution state tracking and checkpoint recovery. |
| Checkpoint | /uc:checkpoint | Saves execution state for session recovery. Triggered automatically every 3 tasks or when a usage soft/limit block is recorded. |
Documentation & Verification
| Skill | Command | Description |
|---|
| Docs Manager | /uc:docs-manager | Guards canonical documentation/ structure. Routes docs to correct directories, keeps each fact in one canonical home linked by standardized anchored cross-links, maintains index. |
| Verification | /uc:doc-code-verification-mode | Compares doc claims against code reality using Checker agents, and detects content duplicated across docs and broken cross-reference anchors. Classifies discrepancies with fix types including consolidation. Extends the planning framework. |
| Context Management | /uc:context-management | Manages context/ directory for external API docs and SDK references with git submodule support. |
Project Management
| Skill | Command | Description |
|---|
| Backlog | /uc:backlog | Four-category backlog (bugs/questions/ideas/debt) with priorities and blocking relationships. User-initiated only. |
| Plan Status Sync | /uc:plan-status-sync | Scans plans, infers status from execution artifacts, reconciles README with plan.json. |
| Help | /uc:help | System advisor: recommends the right skill, agent, or workflow for any task. Shows recent changelog. |
Infrastructure
| Skill | Command | Description |
|---|
| Harness Builder | /uc:harness-builder | Knowledge base for building harness components (skills, agents, hooks, protocols): description budgets, a measured structural-duplication catalogue, efficient communication via the dormant caveman engine, session context-cost auditing, before/after refactor testing, agent-building knowledge. Build tasks run a mandatory staged workflow editing in place on a clean git state — structural (user-confirmed), lexical (user-confirmed), description & name (user-confirmed), engine compression via the Caveman Compress agent consumed item by item, then git diff presentation and explicit approval before commit; knowledge questions answer directly, and active UC planning modes take precedence. |
| Rename Window | /uc:rename-window | Renames the current tmux window. Applies the UC::P-NNN::<plan> / UC::<Mode>::<subject> convention that planning modes and plan-execution set automatically (plan ID takes priority). |
| Railway | /uc:railway | Manages Railway.com deployments via CLI with multi-account token switching and config-as-code. |
| Tailscale Setup | /uc:tailscale-setup | Exposes local services via Tailscale serve/funnel with HTTPS wrapping. |
| Update | /uc:update | Updates Ultra Claude via marketplace, restarts layout daemon, checks for pending migrations. |
Reference Libraries
Reference libraries are shared instruction sets — not skills. Planning modes inherit them and extend them per-stage via files in their own references/ directories.
| Library | Path | Description |
|---|
| Planning Framework |
references/planning-framework/ |
4-stage planning flow (Understand → Research → Discuss → Write), conversational rules, existing-plan handling, approval gates, and post-approval hard stop. Inherited by feature-mode, debug-mode, and doc-code-verification-mode via per-stage extensions in each mode's references/stage-N.md. |
Agents
Agents are spawned by skills via the Agent tool — execution pipeline roles as persistent teammates (named, background) that coordinate through the full task lifecycle, and research/survey roles as stateless one-shot subagents (unnamed, sync or fan-out) that return a result and exit.
| Agent | Model | Role |
|---|
| Task Executor | Per task: Sonnet / Opus (default) / Fable | Per-task team hub: explores codebase, writes plan, implements code, drives review/test cycles. Runs solo on ops tasks (deploy/verify/monitor) |
| Code Reviewer | Sonnet | Read-only quality gate: reads standards early, gives plan feedback, formal pass/fail review |
| Task Tester | Sonnet | Sends upfront TESTER TAKE, authors black-box acceptance tests, verifies against product docs, checks UI in browser |
| Project Manager | Sonnet | Event-driven coordinator: maintains execution state files, reads signals.jsonl per task for stage tracking, per-task budget tracking, operational reports. Owns the background liveness monitor and verifies its NUDGE candidates, pinging wrongly-parked agents and escalating to Lead only on confirmed non-response. Does not monitor usage — it consumes the limit sentinel's usage events passively for budget tracking. Wakes only on messages or actionable monitor emits. |
| Fresh Eyes | Fable | Fresh-context review of finished written work — a plan directory, a doc set, any scope. Offered once at Stage 4 of every planning mode, right after the plan is written and before the approval discussion; also spawnable ad hoc. Reads the work and the docs and code it stands on with none of the conversation that produced it, and reports wrong solution, unforced complexity, approaches out of step with the field, performance risk, unmet business requirements. Propositions only — never edits, never gates. |
| Researcher | Sonnet | Stateless one-shot researcher spawned by the /uc:research skill on cache miss. Fetches library/pattern/market docs via Ref.tools or web search, writes to documentation/technology/research/ or documentation/product/research/, atomically updates the index. Never reads project source code. |
| Checker | Sonnet | Single-topic doc-vs-code comparison with severity levels and file:line references |
| Code Surveyor | Sonnet | Fast structural code scan: catalogs files, components, patterns for orchestrators |
| Doc Surveyor | Sonnet | Documentation section exploration: identifies content types, topics, specs |
| System Tester | Sonnet | Scientific bug reproduction: follows exact steps, tries variations, never fixes code |
| Caveman Compress | Opus | Compression engine wrapper: runs the caveman-compress CLI on a scratch copy of a body artifact (prompt bodies, doc sections, protocol formats — never descriptions, whose frontmatter the engine preserves verbatim) and returns the compressed version plus a cut list tagged clean / fixable (with repaired wording) / harmful — proposition-only, never edits outside scratch, no yield threshold. Spawned at stage 4 of /uc:harness-builder's build workflow; the parent adopts cuts item by item. |
Not an agent: the limit sentinel. Usage limits are handled by one machine-global background process (scripts/limit-sentinel.sh, installed by /uc:setup), not by any agent. It detects limit hits via the StopFailure hook, wakes parked sessions at each window reset (+90s), injects 90% soft-band advisories into Lead panes, wakes a plan whose window ended in the soft band even though nothing was ever limit-killed (the fleet stopped itself, so nothing is parked), fires the same wake at the weekly (7d) reset, keeps a 5h usage window open continuously via a low-cost heartbeat so windows tile back-to-back, and notifies you when a weekly limit is hit.