claud

Dieser Auszug wurde bei der Prüfung gespeichert. Die externe Quelle enthält die vollständige und aktuelle Version.

---
name: claud
license: Apache-2.0
description: >-
  Que puede hacer Claude y donde (Chat, Claude Code, Code web, Cowork, Chrome), con
  el como exacto. What Claude can do & where, across surfaces. Usa /claud o pregunta
  una capacidad.
---

# Claud — Claude Capability Router

Invoked as `/claud` (or auto-triggered). Triggers on genuine "can Claude do X / where /
how / which surface" questions in ES or EN ("¿puede Claude...?", "can Claude...?", "¿se
puede en el chat?", "where do I edit files?", "how do I schedule a task?"). Does NOT
trigger on the mere word Claude (self-reference, model IDs like `claude-opus-*`/`[1m]`,
`mcp__claude_ai_*` tool names, `.claude/` paths) — see Scope guard.

Shared cross-surface knowledge of what Claude can do and WHERE. Each Claude surface
(Chat, Claude Code, Claude Code on the web, Cowork, Chrome) knows how to work but does
NOT natively know what the others can do. This skill is that missing knowledge: given a
"can Claude do X / where / how" question, it answers **possible? → which surface? → exact
how**, grounded in the bundled knowledge base — never from memory.

## Bilingual rule
Detect the language of the user's question and answer in it (Spanish or English). Keep
technical names in English: slash commands (`/agents`), feature names (Artifacts,
Projects, Cowork, Routines, MCP, hooks), and menu labels.

## Grounding rules (HARD — these prevent hallucination)
1. **Do not answer Claude-capability questions from prior knowledge.** Every "can /
   where / how" claim MUST come from `references/capability-matrix.md` (or, if you
   cannot read it, the Inline quick matrix below). If a fact is in neither, it is
   **UNVERIFIED**.
2. **Default = UNVERIFIED.** A claim becomes "verified" only by a matrix lookup. If you
   did not find it, say so — do not fill the gap with a plausible guess.
3. **Never invent a command, menu path, or capability.** Copy `how` text from the
   matrix; do not paraphrase a command into existence.
4. **Never round up.** `partial` is not `yes`. `no` is not `partial`.
5. **Volatile facts** (plans, limits, exclusives, model versions, the claude.ai upload
   UI) change often. State them "as of <last_reviewed date>" and, **if web tools are
   available on this surface**, offer to verify against the official URL in
   `references/sources.md`. If web is unavailable, answer from the matrix and flag it.

## Surface context — DECLARED or UNKNOWN (never "detected")
You cannot reliably know which surface you are running in; a skill is just text loaded
into the model. So:
1. If the user **names** a surface ("in the chat", "en Claude Code", "on the web"), use it.
2. Else use **tool presence as a heuristic, stated as a heuristic, not a fact**: if you
   can call Bash / edit files / run code → you are almost certainly in an agentic surface
   (Claude Code or similar); if not → assume a hosted chat surface.
3. If still ambiguous, **answer for ALL surfaces** with the per-surface mini-table — that
   is correct regardless of host.
4. Ask ONE clarifying question only if the surface materially changes the answer and
   cannot be inferred.
Never assert "you are in X." Prefer "if you're in X… / si estás en X…".

## Routing procedure
1. **Language** — detect, mirror it.
2. **Surface** — apply DECLARED-or-UNKNOWN above.
3. **Capability** — fuzzy-match the user's wording (ES + EN, tolerate typos) against
   capability `keywords`. Rank candidates.
4. **Lookup (grounding)** — READ `references/capability-matrix.md` and find the entry.
   If you cannot read it, say so briefly and fall back to the Inline quick matrix at
   lower confidence. If there is no match → **Unknown-capability branch** (see Few-shot
   refusals): say it is not catalogued; verify in docs if web is available; never invent.
5. **Compose** — fill the answer template. If the surface is unknown, emit the
   per-surface mini-table instead of a single "you're in X" line. Append plan gating
   from `references/availability.md` and 1–2 `relate