Skip to content

create-workflow.md

Latest commit

 

History

History
51 lines (36 loc) · 4.35 KB

create-workflow.md

File metadata and controls

51 lines (36 loc) · 4.35 KB
Raw
description Scaffold a new agent-relay multi-agent workflow using the writing-agent-relay-workflows skill
argument-hint
task description
allowed-tools Read, Write, Edit, Bash, Glob, Grep

Create Agent-Relay Workflow

Scaffold a new multi-agent workflow that runs on the agent-relay broker. This command is harness-agnostic — works under Claude Code, Codex, OpenCode, Droid, Gemini, or any harness that supports markdown slash commands.

Task: $ARGUMENTS

Instructions

  1. Load the skill. Read the writing-agent-relay-workflows skill. After installation with npx prpm install @agent-relay/writing-agent-relay-workflows, look in the harness-managed skill locations first: .claude/skills/writing-agent-relay-workflows/SKILL.md or .agents/skills/writing-agent-relay-workflows/SKILL.md. When developing inside this repo, the same source lives at skills/writing-agent-relay-workflows/SKILL.md. Do not proceed without it — the SDK surface, DAG semantics, and verify-gate patterns it documents are required.

  2. Pick the orchestration pattern. Consult the choosing-swarm-patterns skill. After installation with npx prpm install @agent-relay/choosing-swarm-patterns, look in .claude/skills/choosing-swarm-patterns/SKILL.md or .agents/skills/choosing-swarm-patterns/SKILL.md; in this repo, read skills/choosing-swarm-patterns/SKILL.md. Use it to select among fan-out, pipeline, hub-spoke, consensus, mesh, handoff, cascade, dag, debate, or hierarchical based on the task shape. State the chosen pattern and why in one sentence before writing code.

  3. Author the workflow. Using the WorkflowBuilder API from the skill:

    • Define each step with explicit inputs and {{steps.X.output}} chaining.
    • Add a verify gate that fails closed on missing evidence.
    • Add the selected review-depth fresh-eyes path: light uses Claude review -> fix; standard adds final Claude review -> final fix; deep adds the full Codex review/fix/final-review/final-fix path after the Claude path. Default to deep when the task is ambiguous, production, security, billing, destructive, or broad multi-file work.
    • Make final acceptance depend on final-review-pass-gate, final hard validation, scoped diff evidence, and regression gates for the selected review-depth path. Never let light or standard skip deterministic proof.
    • Keep prompts model-agnostic — never hardcode a specific model name into a step's instructions.
    • Size steps so a single agent can complete them in one focused pass.

  4. Wire the team. Define the lead + worker agents, the channel(s) they share, and the team name. Reuse names from agent-relay agents only after checking for collisions.

  5. Provide a runnable example. Output one minimal end-to-end example that demonstrates feeding $ARGUMENTS into the workflow and shows the expected {{steps.*.output}} shape at each stage.

  6. Verify the agent-relay CLI is available. Before recommending any dry-run or launch command, confirm the runtime that will execute the workflow is installed:

    command -v agent-relay || npx agent-relay --version

    If neither resolves, note that it must be installed (npm install -g agent-relay, or invoke via npx agent-relay …) and surface that in the integration notes rather than assuming it is present.

  7. Integration notes. End with a short checklist for wiring into the existing workload-router (where the workflow file lives, how it's registered, how to dry-run, how to launch via agent-relay).

Output Contract

  • The workflow source file (TypeScript or YAML — match the surrounding repo convention).
  • A one-paragraph summary: pattern chosen, agent roster, verify gate, selected review depth, and review/fix path.
  • Integration checklist (5 bullets max).

Constraints

  • Model-agnostic prompts only. No "use claude-opus" or "use gpt-5" inside step instructions.
  • Verify gates must check evidence (artifacts, test results, file contents), not self-reports.
  • Review fix steps must harden fixes with appropriate tests, fixtures, assertions, or deterministic proof commands whenever the finding is testable.
  • Final acceptance, commit, PR creation, or handoff must depend on the selected review-depth path plus final deterministic gates, not directly on implementation or an informal lead review.
  • Do not invent SDK APIs — if the skill doesn't document it, ask before adding.