Skip to content
Agent Orchestration
Skill

/team-routes

dev-orchestrator only. Turns any request into a route of specialist and writer steps: which roles work, in what order, what each reads and writes. Use at the start of every new goal. SKIP: writer CLIs; specialists themselves.

BOOST
From plugin
claude-lane-stack
12464 skills24 agents5 commands12 hooks
+1
Install
$ npx -y skills add VKirill/claude-lane-stack --skill team-routes --agent claude-code

How it fires

How this skill gets triggered: by you, by Claude, or both.

  • Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
  • You can call itInvoke it directly when you want it.
  • Slash command/team-routes

Context preview

The summary Claude sees to decide when to auto-load this skill.

dev-orchestrator only. Turns any request into a route of specialist and writer steps: which roles work, in what order, what each reads and writes. Use at the start of every new goal. SKIP: writer CLIs; specialists themselves.

SKILL.md

team-routes.SKILL.md
name: team-routes
description: "dev-orchestrator only. Turns any request into a route of specialist and writer steps: which roles work, in what order, what each reads and writes. Use at the start of every new goal. SKIP: writer CLIs; specialists themselves."

Team routes

The operator states a goal, not a staffing plan. You decide who works on it. A page built without its SEO, UX and copy inputs comes back as placeholders and gets redone; a one-line fix sent through four specialists burns an hour of expensive models. Pick the smallest team that produces the artifact the operator actually needs.

1. Name the deliverable

Before Score/Decompose, write one line for yourself: `deliverable: <artifact> · surface: <page/app/feature> · slug: <kebab-slug>`.

The deliverable is what the operator will open at the end: a gray prototype, a branded mockup, shipped code, an article, an audit report, an answer with sources. Ambiguous wording («сделай страницу») means the cheapest deliverable that still moves the goal (prototype before code) unless the operator said «в код», «на сайт», «задеплой» or the page already exists in code.

2. Pick or compose a route

Start from the closest recipe in [references/recipes.md](references/recipes.md). Drop a step when its output already exists on disk and is not stale (for example `.agents/copy/pages/<slug>.md` is there, so copy only fills gaps). Add a step when the goal needs a role the recipe lacks. No recipe fits → compose a route from the role cards in [references/roles.md](references/roles.md): each step's inputs must be outputs of earlier steps or existing files.

Run steps in parallel when neither reads the other's output. Run them in order when one decides what the next one writes about (SEO intent before copy).

Not every request needs a route. Go straight to the existing Loop when the goal is product code with no new user-facing text or layout (bug, refactor, API, infra), or a question you can answer from the repo.

3. Announce and start

Post one Russian line, then start the first step in the same turn. Do not wait for approval: the operator asked for the result, and the line exists so they can redirect you.

Маршрут: seo ∥ design-lead (UX) → copy-lead → design-lead (прототип) → design-lead (аудит). Начинаю.

For routes with 2+ specialist steps, also write the route file `.agents/plans/routes/<slug>.md` (English):

goal: <operator goal, one line>
deliverable: <artifact>
steps:
  - id: seo | role: seo-specialist | reads: [...] | writes: <path> | status: todo
  - ...
decisions: []   # conflicts you resolved, with the reason

Update `status` after each step. It is how `resume-project` and a restarted session know where the route stopped.

4. Brief each specialist

Every specialist prompt has the same five lines. The specialist sees only this prompt and the disk, so a missing READ path means the work is done blind.

GOAL: <operator goal + this step's part of it>
SLUG: <slug>
READ: <outputs of earlier steps and existing project files, as paths>
WRITE: <the one path this step owns>
DECIDE: <the questions this step must answer for the next step>

Close with the role's normal sentinel (`DONE <path>` / `FAILED <reason>`). Read the file from `DONE`, not the chat summary, before you start the next step.

5. Merge the outputs

Specialists do not talk to each other. You merge their outputs. When two outputs conflict (SEO wants a long FAQ block, design wants one short screen), decide by the deliverable's primary user action, record the call in `decisions:` with one reason, and pass it to the next step through READ/DECIDE.

Ask the operator only for what no role can produce from the disk or the web: prices, guarantees, the offer, legal claims, brand approval, money or irreversible actions. Specialists must mark unknown facts `[unknown]`, not invent them. Collect those markers into one question for the operator at the end of the route, not after each step.

6. Hand over

End the route with the deliverable path, what each role contributed (one line each), open `[unknown]` facts, and the next sensible step (for example «в код — скажи "делай"»). A route that ends in code continues into the normal WRITE Loop: specialist steps finish first, then `run-init` (Lane Pilot: `lane_pilot_dispatch_writer`, specialists via `lane_pilot_specialist`) with their outputs in `read_first`. This keeps the Mode XOR rule: no research teammates on a goal once its write run has started.

Read more
Ships withclaude-lane-stack

Claude Lane Stack is a self-hosted toolkit for coordinating coding agents through CLI processes, task contracts, Git workspaces, and verification receipts (bin/run-controller:117-159, schemas/task-v2.schema.json:1-60).

Get the whole plugin

Other skills on claude-lane-stack.