Nox-Lumen MfgNox-Lumen Mfg

skill-architect

Capabilities

Guides users to design and build high-quality Agent skills from scratch, covering standalone skills and MCP-enhanced workflows.

Typical triggers:

  • “I want to create a new skill”
  • “Help me turn this workflow into a skill”
  • “I want to automate this workflow”
  • “How do I teach my Agent to do X”
  • When you need to author a SKILL.md

Not for: creating subagents or writing technical design documents.

Five-stage workflow

Execute strictly in order; do not skip steps:

Rendering diagram…

① Discovery — clarify the need

  • Understand the problem
  • Define 2–3 concrete examples
  • Clarify triggers and success criteria
  • Identify edge cases and existing tooling

② Architecture — pick pattern + outline structure

Choose a suitable pattern:

PatternWhen to use
Sequential workflowFixed steps, clear order
Iterative refinementMulti-round feedback, convergence
Context-awareOutput heavily depends on current context
Domain intelligenceHeavy domain expertise, LLM-led logic

Plan directory layout and craft description (trigger phrases + negative boundaries).

③ Craft — author content

  • Frontmatter (name in kebab-case + description ≤ 1024 chars)
  • Instruction body (≤ 500 lines)
  • Real examples
  • Error-handling paths

④ Validate — check quality

  • Structure (frontmatter format, kebab-case, no README.md as substitute)
  • Trigger test (does the description actually invoke correctly?)
  • Instruction review (ambiguity, redundancy, gaps)

⑤ Deliver — ship

Use skills_create(name, description, content). Pass only Markdown body (# heading onward); no --- frontmatter block in content.

Hard rules

RuleRequirement
namekebab-case, no spaces, no uppercase
description[Capability] + [trigger phrases] + [out of scope], ≤ 1024 chars, no XML angle brackets
SKILL.mdBody ≤ 500 lines
Never skip DiscoveryDo not jump straight to SKILL.md

Anti-patterns (avoid)

  • description reads like prose with no trigger phrases
  • Starting Craft before Discovery
  • Pasting raw user chatter into the skill instead of distilled rules
  • Missing negative boundaries (“Not for…”)

How to trigger

/skill-architect I want to turn our “weekly report cleanup” routine into a skill

Natural language examples:

  • “Help me wrap my common SQL queries in a skill”
  • “Teach my Agent how to produce A/B test reports”
  • “Turn this workflow into a skill”

On this page