Custom Skills Tutorial
Skills are focused knowledge bundles that teach an agent when and how to apply a
specialized workflow. LumenFlow’s canonical skill directory is
.lumenflow/skills/.
Create a skill
Section titled “Create a skill”-
Create a directory using a stable, descriptive name:
-
Add
.lumenflow/skills/api-contract-review/SKILL.md: -
Review the skill for portability.
-
Load it through the capability available in your current agent host before the work it governs.
Portability checklist
Section titled “Portability checklist”A canonical skill should:
- describe triggers and outcomes in plain language;
- use repository-relative paths;
- call LumenFlow CLI commands only for LumenFlow behavior;
- refer to delegation, file reading, or web research as host-native capabilities;
- avoid product filenames, proprietary slash commands, editor settings, and assumptions about a particular model;
- state any required external capability explicitly and provide a graceful fallback where possible.
Transposition
Section titled “Transposition”Prefer a host that can read the canonical SKILL.md directly. If another
location or wrapper is required, the repository owns that transposition:
- wrappers should read the canonical file at invocation time;
- discovery adapters should include the canonical content digest and fail closed when the source is unavailable;
- symbolic links are a repository portability choice;
- copied skills are snapshots and must be updated by the repository owner.
LumenFlow install, upgrade, docs-sync, force, merge, and uninstall never modify the transposed artifact.
Reserved and project skills
Section titled “Reserved and project skills”LumenFlow refreshes only its reserved skill allowlist. A project-created skill
with another name is left untouched. Use a project namespace when collision
risk matters, and keep the skill’s name stable once other guidance refers to
it.
The reserved browser-resource-etiquette skill applies whenever work uses a
browser, preview, or browser-backed verification. It requires an isolated
session and limits cleanup to browser resources owned by that session, so an
agent never closes a user’s browser or another task’s processes.
Test a skill
Section titled “Test a skill”Test the actual outcomes the skill promises:
- give an agent a small representative task;
- confirm the trigger description makes selection clear;
- verify required sources and commands exist;
- check that the result follows the workflow without depending on one host;
- run repository docs and formatting gates.
For the universal ownership boundary, see Connect an AI Coding Assistant.