Przejdź do treści
Biblioteka skilli
AI i agenci · Planowanie

skill-router

/skill-router

Kieruje zadanie do właściwego skilla, agenta lub narzędzia MCP na podstawie opisu potrzeby, bez jego wykonywania.

Repozytorium
agentbox ↗
Licencja
AGPL-3.0
Ostatnia zmiana repozytorium
27.10.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
4
Skrypty
1
Komentarze HTML
0
Adresy URL w SKILL.md
0
Przestarzałość

Brak sygnałów. Repozytorium zmienione 27.10.2026, bez odwołań do wycofanych modeli i API.

Licencja

AGPL-3.0, copyleft: przeróbki udostępniane innym muszą mieć tę samą licencję.

Co sprawdzić przed instalacją
  • 1 skrypt w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.

Skrypty: scripts/route.mjs

Co robi i kiedy

Kieruje zadanie do właściwego skilla, agenta lub narzędzia MCP na podstawie opisu potrzeby, bez jego wykonywania.

Oryginalny opis (EN)

Unified dispatcher for the full skills estate. Use when you don't know which skill to invoke — describe your task and get routed to the optimal skill, agent composition, or MCP tool. Not for executing the task itself: it only routes, then hands off to the skill that does the work.

Jak działa

  1. Usage
  2. Routing Method
  3. Routing Rules
  4. What This Skill Does NOT Do
  5. Maintaining the routing table
  6. Cross-harness note
Sekcje z nagłówków SKILL.md, w oryginale · 107 linii, 781 słów
Podgląd SKILL.md
---
name: skill-router
description: >
  Unified dispatcher for the full skills estate. Use when you don't know which skill to invoke —
  describe your task and get routed to the optimal skill, agent composition, or MCP tool. Not for
  executing the task itself: it only routes, then hands off to the skill that does the work.
version: 1.1.0
author: agentbox-claude
tags:
  - routing
  - discovery
  - dispatcher
  - meta-skill
user-invocable: true
status: router-only
---

# /route — Unified Skill Dispatcher

Describe your task. Get routed to the right skill. You don't need to know the full skills estate — just say what you need.

## Usage

```
/route [describe what you need]
```

Examples:
- `/route fix the login bug and add tests`
- `/route generate a Wardley map of our infrastructure`
- `/route audit this UI for accessibility`
- `/route research competitor pricing for UK market`
- `/route harden this Linux server for SOC2`
- `/route 123 Main St, Brooklyn — site analysis and zoning`
- `/route make a podcast about our architecture`

## Routing Method

Two paths, chosen by `[skills.routing].router` in `agentbox.toml` (ADR-2091). The default
is the judge; the table is its fallback and is never removed.

**1. Ask the judge (default, `router = "jev"`).** Run, with the user's input verbatim:

```bash
node scripts/route.mjs "<everything after /route>"
```

It puts the request to TypeSafe System One as one Choice over every routable skill's own
description (the `system-one` skill owns that backend) and prints ranked picks with
probabilities and a `dispatch:` line — ~1 s, ~15k input tokens, about $0.0006. Take the
`dispatch:` skill to the routing rules below. The probability is advisory: a wrong pick has
been measured at 0.94, so a low number is a reason to ask, and a high number is not proof
(ADR-2090). `dispatch: none` means the judge saw no skill that would change your approach —
answer directly. `--json` gives the raw outcome; `--router table` forces path 2.

**2. Read the table (fallback, or `router = "table"`).** Whenever the script prints
`router: table (fallback: …)` — judge off, no key, timeout, 429/529, any error — do what
`/route` always did:

1. Load **`references/routing-table.md`** (generated from every skill's front-matter
   description and `references/section-map.json`; not inlined here to avoid drift).
2. Classify intent against its sections (they mirror the category headings of
   `SKILL-DIRECTORY.md`; one row per skill, each row's text is the skill's trigger description).
3. Apply the routing rules below to dispatch, clarify, or compose.

The fallback is the normal path when the judge is slow, not an error to report — say
`(routed by table)` in your one-sentence dispatch and carry on.

The same library also runs as a per-turn hook (`config/hooks/skill-route.cjs`) that injects
one short line — `[route] x 0.91 · y 0.05 · z 0.02 — advisory; load a skill only if it fits
this turn.` — into ordinary turns; `/route` is for when you want the full ranking or the hook
is off.

## Routing Rules

### Rule 1: Clear match → dispatch immediately
State which skill handles the request in one sentence, then invoke it.

### Rule 2: Ambiguous → ask ONE question
If the intent could go to 2+ skills, ask exactly one clarifying question. Then route.

### Rule 3: Multi-skill composition → state the plan
If the task spans multiple skills, state the sequence and invoke the first one.
Example: "Starting with `perplexity-research` for competitor data, then `report-builder` for the analysis document."

### Rule 4: No match → show condensed menu
```
I don't have a specific skill for that. Here's what I cover:

• Code: /route [bug fix / feature / refactor]
• Research: /route [web search / URL analysis / NotebookLM]
• Economics: /route [GPU cost / agent job pricing / token conversion / valuation]
• Design: /route [UI/UX / design audit / typography]
• Docs: /route [report / LaTeX / diagrams / Wardley map]
• Media: /route [image / video / 3D / AI art]
• DevOps: /route [GitHub / CI-CD / release]
• Security: /route [hardening / compliance / audit]
• Architecture: /route [review / first-principles / entropy lens]
• AEC: /route [site planning / zoning / sustainability]

Or browse the full inventory: see SKILL-DIRECTORY.md
```

### Rule 5: Just `/route` with no arguments → show condensed menu

## What This Skill Does NOT Do
- It does not execute tasks. It routes to the skill that does.
- It does not override skill-internal logic.
- It does not ask more than one clarifying question.

## Maintaining the routing table
`references/routing-table.md` is **generated**: `node skills/gen-routing-table.mjs` reads every
skill's frontmatter `description` (and `deprecated`/`replacement`) and the section for each
skill from `references/section-map.json`, and rewrites the table. `bash skills/lint-skills.sh`
fails when the table is stale (`--check`) or a skill directory is missing from the section map.
Never hand-edit a row: fix the owning skill's description, add the skill to the section map if
it is new, and regenerate.

## Cross-harness note
`/route` is a Claude Code slash command. On Codex run `scripts/route.mjs`
the same way (it is plain Node, no harness affordance) or read
`references/routing-table.md` directly; the same rows, the same descriptions.
Pliki w paczce (4)
  • SKILL.md 5 KB
  • references/routing-table.md 62 KB
  • references/section-map.json 8 KB
  • scripts/route.mjs 6 KB

Wersje

RepozytoriumWynikLicencja
Ta strona · dreamlab-ai/agentbox100AGPL-3.0
FrancoStino/opencode-skills-collection100MIT
hanmariyang/giting-skills100MIT
wrongstack/wrongstack100MIT
sickn33/agentic-awesome-skills100MIT
dotslashsimran/simcraft100MIT
borghei/Claude-Skills100MIT
byrmeng/skill-router100MIT
aminmzadev/keyless-agent-skills100MIT

Migawka bazy z 13.11.2026 · wersja paczki af4b7eeca562 (skrót SHA-256 całej paczki)

Pobrania w ostatnich 30 dniach: 0 · ranking

Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.

Podobne skille

  • 86ctxora-navigation
    Kieruje zadanie CTXORA do najmniejszego właściwego skilla, polecenia, roli agenta i zestawu narzędzi MCP.
  • 90team-router
    Kieruje zadanie do właściwego poziomu wspólnego banku skilli i zestawu ról oraz podaje, które skille zainstalować.
  • 96skills-gateway
    Kieruje zadania do właściwego skilla z katalogu niezarejestrowanych skilli, np. tekstu, JSON albo git, i ładuje jego instrukcje.
  • 100reasoning-router
    Klasyfikuje trudne problemy i kieruje je do właściwego z jedenastu serwerów myślenia MCP, a przy spadku pewności przekierowuje zadanie.
  • 100agent-router
    Kieruje zdarzenia w systemie wielu agentów do właściwych workerów i zbiera wyniki w jedną odpowiedź, bez wykonywania pracy samemu.
  • 100task-intake-router
    Kieruje przychodzące żądanie do właściwego trybu, typu agenta, poziomu modelu i wzorca delegowania, zanim zacznie się implementacja.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.