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__claudeai 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.

1. 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.

1. Never invent a command, menu path, or capability. Copy how text from the

matrix; do not paraphrase a command into existence.

1. Never round up. partial is not yes. no is not partial.
2. 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.

1. If still ambiguous, answer for ALL surfaces with the per-surface mini-table — that

is correct regardless of host.

1. 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.

1. 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.

1. 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 related capabilities if useful.

1. Disambiguate — only if 2+ candidates diverge into different answers, ask ONE

question with the options.

Inline quick matrix (degraded-mode fallback)

Use this only if you cannot read references/capability-matrix.md. ✅ yes · ⚠️ partial · ❌ no. Surfaces: Chat | Code (CLI/IDE) | Web (Claude Code on the web) | Cowork (Desktop) | Chrome. As of 2026-06-14.

Answer template (bilingual)

Skip empty slots. Keep it scannable.

[Capability] — YES / NO / PARTIAL   (SÍ / NO / PARCIAL)
✅ <where it works> → <terse how>
⚠️ <where partial> → <caveat>
❌ <where it doesn't> → <one-line why>
👉 Recommended: <surface> — <exact command / menu / steps>
   (if the user's surface is UNKNOWN, give the per-surface mini-table instead)
📍 If you're in <surface> / Si estás en <surface>: <"you can, do X" | "not here, go to Y">
   (only when the surface was declared or strongly inferred; omit if unknown)
💳 Plan: <gate>            (only if plan-gated)
🔗 Also / También: <1–2 related capabilities>

The three core questions (possible? / where? / how?) map onto this template; the extra slots just make the "where/how" precise. Skip any slot that doesn't apply.

Example (ES) — usuario en el Chat: "¿puedo programar una tarea recurrente?"

Programar una tarea recurrente — PARCIAL (depende de la superficie)
✅ Claude Code on the web → Routines (cron/presets); corre en la nube aunque cierres la laptop.
✅ Claude Cowork → tarea recurrente local; SOLO corre mientras Claude Desktop esté abierto.
⚠️ Claude Code (CLI) → `/schedule` crea una Routine en la nube.
❌ Chat y Chrome → no ejecutan nada por sí solos en un horario.
👉 Recomendado: si es código sobre un repo → Routines (claude.ai/code). Si es trabajo de
   oficina (reportes, monitoreo) → Cowork.
📍 Si estás en el Chat: aquí no se puede; cambia a Cowork o a Claude Code on the web.
💳 Plan: Routines y Cowork requieren plan de pago.

Example (EN) — "can Claude send my emails?"

Send email — NO (on every surface)
❌ The Gmail connector can search, read, and create DRAFTS, but Claude cannot SEND on your behalf.
👉 Recommended: let Claude draft it (Chat/Desktop with the Gmail connector), then you press Send.
🔗 Also: manage calendar events (create/update/delete) — that one Claude can do directly.

When to read references

• references/capability-matrix.md — READ FIRST for almost any capability question

(the single source of truth; per-surface how lives only here). Fallback: Inline matrix.

• references/surfaces.md — surface disambiguation + per-surface auth/navigation/limits.
• references/availability.md — plan gating (Free/Pro/Max/Team/Enterprise) and exclusives.
• references/sources.md — official doc URLs + last-verified dates (for web verification).

If a reference cannot be read on this surface, say so and answer from the Inline matrix, flagged at lower confidence.

Few-shot refusals (model these when grounding fails)

• Not catalogued: "Eso no lo tengo catalogado en mi base de capacidades, así que no

puedo confirmarlo. Si quieres, lo verifico en la documentación oficial." / "That isn't in my capability reference, so I can't confirm it — I can check the official docs."

• Reference not loaded: "No pude abrir mi matriz de capacidades en esta superficie;

respondo desde la tabla rápida con menor confianza: …" / "I couldn't open my capability matrix here, so this is from the quick table at lower confidence: …"

• Stale/volatile: "Esto cambia seguido. Al 2026-06-14: … ¿Verifico la versión actual

en docs?" / "This changes often. As of 2026-06-14: … Want me to verify in the docs?"

Scope guard

If the user just wants the task DONE and the current surface can do it, do it — don't deliver a lecture about surfaces. Route only when they ask what's possible / where / how, or when the current surface cannot do what they want.

Do NOT engage the router when the word "Claude" appears in a non-capability context: self-reference, model IDs (claude-opus-, [1m]), mcp__claudeai tool names, or .claude/ paths. In those cases answer normally — no capability table. If a message both mentions Claude AND asks a genuine can/where/how question, route only the latter part.