| name | port-skill-creator |
| description | Author a new Agent Skill for this repo (port-labs/port-skills): scaffold the skills/port-<name>/ directory, write valid frontmatter, structure the body as a short router into references/ and assets/, validate it, and regenerate the README index. Use when asked to create a new Port skill, add a skill to this repo, help me write a SKILL.md, contribute a skill, port an internal Port AI skill to this repo, or fix a skill that fails validate-skill.js. |
| license | MIT |
| compatibility | Claude Code, Cursor, Codex CLI, GitHub Copilot |
| metadata | {"version":"2.0.0","author":"port-labs","repository":"https://github.com/port-labs/port-skills","tags":"port,meta,skill-authoring,reference","summary":"Author and validate a new skill for this repo"} |
Skill creator
A skill here is a plain file in this git repo, read by a third party's coding
agent, that helps it get a Port job done: build a context lake, write a
mapping, manage Port with Terraform. It is not Port's internal in-product
AI skill system (packages/ai-skills in port-labs/port, which creates a
skill entity via upsert_entity for Port AI's own runtime). See
references/skill-writing-guide.md#adapting-an-internal-skill
if you're starting from one of those.
This skill itself lives at .claude/skills/port-skill-creator/, not under
skills/, because it's for someone contributing to this repo, not for
someone who installed skills/ into their own coding agent to work with
their Port account. It never ships to end users; it's only relevant when
working in this repo, so it stays repo-local.
Prerequisites
None. This is a reference skill, no live Port account or MCP server needed,
just this repo and Node.js to run the scripts below.
File layout
skills/port-<name>/
├── SKILL.md # required: short frontmatter + router body
├── references/ # one file per sub-topic, the actual depth
└── assets/ # images, scripts, and complete ready-to-run examples
<name> must be lowercase, hyphenated, prefixed with port-, and match the
frontmatter name exactly (enforced by validate-skill.js).
How to build a skill
- Scope it. One sentence on what job this helps do, what it assumes
already exists, and what's explicitly out of scope (point to the skill
that owns that instead). Check the README so scopes
don't overlap. See references/skill-writing-guide.md#scope-it-first.
- Choose reference vs. MCP-powered. See
references/skill-writing-guide.md#choose-a-skill-class.
- Write the frontmatter, description last.
description is the only
place "when to use this" lives, packed with realistic trigger phrases.
Never restate it as a body section. See
references/skill-writing-guide.md#description-the-trigger-signal.
- Write the body as a router, not a manual. Assume the agent already
decided to do the job. State prerequisites, then a short numbered list of
what to do, each step pointing at the
references/*.md file that has the
actual depth. Target well under 100 lines. See
references/skill-writing-guide.md#body-is-a-router
for the shape and a worked example.
Prerequisites should point to the port-getting-started skill for account
setup and MCP connection instead of restating it, and should say
explicitly that this skill can use Port's MCP server to apply the work
directly when connected, and that search_port_knowledge_sources covers
anything the skill itself doesn't. See
references/skill-writing-guide.md#prerequisites-port-getting-started-and-mcp.
- Push real content into references/ and assets/. One reference file
per sub-topic (
references/blueprints.md, references/mirror-properties.md,
not one giant references/details.md). Assets are anything ready to use
as-is: images, scripts, complete examples.
- Validate:
node .claude/skills/port-skill-creator/scripts/validate-skill.js skills/port-<name> (or --all).
- Regenerate the README index: .
Common pitfalls
| Symptom | Cause | Fix |
|---|
validate-skill.js reports "Unexpected fields in frontmatter" | A field outside name, description, license, allowed-tools, metadata, compatibility | Remove it, or move it under metadata |
| "Directory name must match skill name" | name: and the skills/<dir> folder differ | Rename one to match the other |
| "Skill name must start with the 'port-' prefix" | Every skill in this repo is namespaced port-* | Rename the directory and name: to port-<name> |
| Skill never triggers in real use | description is generic instead of listing trigger phrases | Rewrite per references/skill-writing-guide.md#description-the-trigger-signal |
SKILL.md reads like a manual, not a router | Real instructions were written inline instead of in references/ | Move the depth out, leave a one-line pointer per step |
| Body has a "Use this skill when" section | Old convention; that's the description's job now | Delete the section, fold anything useful into description |
| README table doesn't show the new skill | generate-skill-index.js wasn't run | Run it, commit the README diff |
| Prerequisites restate account signup and MCP setup inline | Should point to port-getting-started instead | Replace with a pointer, keep only what's specific to this skill |
Quick reference
- Layout:
skills/port-<name>/{SKILL.md, references/, assets/}.
- Frontmatter allowed fields:
name, description, license,
allowed-tools, metadata, compatibility.
- Two classes: reference (no live account needed) and MCP-powered
(uses Port MCP tools when connected, documented fallback when not).
- Body is a router: intro,
Prerequisites, a short numbered list of steps
that each point into references/, nothing more. No Use this skill when section, that's what description is for.
- Prerequisites point to
port-getting-started for setup, and note that Port's
MCP server (when connected) can apply the work directly, with
search_port_knowledge_sources for anything uncovered.
- Validate:
node .claude/skills/port-skill-creator/scripts/validate-skill.js skills/port-<name> (or --all).
- Regenerate the README index:
node .claude/skills/port-skill-creator/scripts/generate-skill-index.js.
- Full authoring guidance, examples, and the skill template:
references/skill-writing-guide.md.
