references/workflow-authoring.md
Workflow Authoring
Use this reference when adding or reviewing workflow skills.
Workflow Checklist
- Name the real user outcome in the first paragraph.
- Define only the blocking onboarding questions.
- Tell the agent what artifacts to gather with Firecrawl.
- Specify the final deliverable shape.
- Include an evidence or citation expectation when claims come from websites.
- Identify work that can run in parallel.
- Keep instructions generic enough for any coding agent harness.
Harness-Agnostic Language
Use:
- "ask the user a clarifying question"
- "use sub-agents if available"
- "run independent page research in parallel"
- "use Firecrawl through the CLI or equivalent Firecrawl tools"
Avoid:
- editor-specific function names
- assumptions about a single modal UI
- hardcoded sub-agent APIs
- output paths that only one harness can access
Lightweight Onboarding
Do not run a long interview by default. First infer from the user's message, files, URLs, and surrounding context. If the agent can safely start, start.
Ask at most 1-3 concise clarifying questions only when a required input is missing or ambiguity would materially change the work. Prefer defaults for non-blocking choices and state them briefly.
Automation Inputs
For recurring jobs, each workflow should be expressible as:
workflow: skill-name
cadence: weekly
inputs:
subject: example
urls: []
competitors: []
output: report.md
The skill does not need to implement a scheduler. It should make the inputs and outputs stable enough for another system to schedule it.
Reference Repos
anthropics/skills: use simple skill directories and minimal metadata.anthropics/financial-services: borrow outcome orientation and optional agent packaging ideas.anthropics/claude-for-legal: borrow vertical workflow organization and concrete deliverable framing.
Do not copy heavier managed-agent scaffolding unless this repo starts shipping standalone hosted agents.