Choose reusable AI methods from the Prompt Library, place a Skill node when a method should be explicit in the graph, edit an owned copy safely, and understand how packages, skill sets, surface gates, and renditions determine the instructions each AI surface receives.
Where: Flowgen → left rail → Prompt Library, or Add → HTML / Code → Skill.
Key ideas
- A skill is a reusable method: A skill packages instructions plus metadata that says when it applies, which surfaces may use it, which tools it requires, and which optional reference or rendition files belong to it.
- The Skill node makes a method explicit: The node has no input. It emits the chosen instructions as typed text from Skill output, so one named method can feed several compatible AI steps.
- Skill sets compose one hop: An explicitly selected or default-on skill can pull named siblings from chain_with. The resolver follows one hop, removes duplicates, blocks cycles, and still applies surface, kind, tool, and prompt-budget gates to every member.
- Renditions adapt one package: A package may include surface- or tool-specific renditions. The resolver chooses the rendition that matches the current surface and may include additional tool-scoped renditions only when their tools are callable.
Steps
- Open Prompt Library from the Flowgen left rail and choose the Skills group.
- Press / when you are not typing to focus search, then search by id, label, description, or trigger phrase.
- Narrow the list with Live only, Family, Surface, Kind, and Rendition filters; the visible/total count confirms how much remains.
- Open a result to inspect its instructions, surfaces, tools, triggers, references, and available renditions before using it.
- Add a Skill node from HTML / Code when the method should be visible, named, wired, or shared across several nodes.
- Pick a built-in or imported skill on the card and review the Writes label so you know whether it governs Structure, Media, Style, Motion, Interactions, Code, or CMS fields.
- Select the node to read the full instructions in the properties sidebar. Choose Edit when a built-in needs customization; Flowgen creates a uniquely named owned copy and leaves the shared built-in unchanged.
- Connect skill_out to each compatible Skill or text input that should use the method, then run the consuming node rather than the Skill node itself.
- For a reusable set, store sibling skill names or record ids in chain_with, or link another visible Skill record from the body. Keep every member scoped to the surfaces and tools it really supports.
- Reopen Prompt Library after changing packages or Brain skills and confirm the expected live package, rendition, and set members are discoverable.
How automatic selection works
- Explicit selection has the highest priority, then default_on, then a matching trigger phrase.
- Only explicit and default-on selections expand chain_with; a trigger match never drags a set behind it.
- Every selected or chained package must pass its surface, kind, and available-tool gates.
- When the combined instructions exceed the surface budget, lower-priority entries are dropped deterministically and the resolver reports why.
Package contents and MCP
- SKILL.md contains required frontmatter and the main method.
- references/*.md files hold supporting material and must be linked from SKILL.md.
- renditions/*.md files hold alternate surface/tool instructions declared in frontmatter.
- The MCP f0_guide tool discovers every package that declares the mcp surface and can return the package body, its references, and its renditions. The MCP host remains the reasoning AI.
Built-in, imported, and Brain skills
- Built-ins ship with Flowgen and remain read-only.
- Imported packages appear after built-ins in the Skill picker.
- A Brain item with type Skill can provide the same package fields, including surfaces, kinds, tools, triggers, default_on, and chain_with.
- A body link such as [[Related skill]] can contribute one valid, visible sibling to the set.
Tips
- Use a Skill node for a method that several graph steps share; keep project facts in Context or Brain instead.
- Keep each skill focused enough that its name, Use when, NOT for, tools, and Writes description predict its behavior without opening the body.
Limitations and important notes
- A skill cannot bypass a missing tool, an incompatible surface/kind, or the prompt budget.
- Editing a built-in creates an owned copy; it does not update the package shipped to everyone.
- A chain is deliberately one hop. A chained member does not recursively expand its own chain_with list.
Troubleshooting
A skill is visible in the library but is not injected into an AI turn.
Check that its surface and kind match, every required tool is available, and the package fits the surface budget. Trigger-selected skills also do not expand sets.
The instructions field is read-only.
The selected skill is built-in. Choose Edit to create a user-owned copy, then edit that copy.
A chain member is missing.
Use the exact skill name or Brain record id, confirm the record is visible in this scope, and verify its own surface/tool gates. The resolver reports unresolved references instead of silently inventing a match.