Pi Skills Selector

Reference Skills with $skill-name anywhere in a user message. Type $ to open Pi's native completion popup, then fuzzy-filter by name. Each result shows its description; selecting one leaves shorthand in the editor. Repeat to reference several Skills.

pi install npm:@ian-pascoe/pi-skills-selector

Requires Pi >=1.1.0 and Node >=22.19.0. There are no settings or extra tools.

Submission

Typed, pasted, and selected references are equivalent. On submission:

Review with $code-review and $ponytail.
→ Review with [$code-review](/skills/code-review/SKILL.md) and [$ponytail](/skills/ponytail/SKILL.md).

Paths come from the current Pi Skill Catalogue. The model receives links, not full instructions, and chooses when to read them. The submitted transcript stores the links rather than a separate raw shorthand message.

Only exact known names in prose expand. Unknown names and expressions such as $HOME stay literal, as do escaped \$skill-name, code, and existing Markdown links. Escape or code-format a literal variable whose name matches a Skill. Multiple and repeated references retain their order; path characters are escaped for Markdown links.

The catalogue follows Pi's discovery, trust, and name precedence, including Skills marked disable-model-invocation because selection is an explicit user request. It is available on the first prompt, refreshes with Pi's resources, and works with enableSkillCommands: false.

Integration

Completion stacks over the current provider, retaining ordinary slash/file completion, keyboard controls, and custom editors. Conversion applies to terminal and RPC prompt input, including steering and follow-ups, without changing attachments or message routing. Extension-generated input is unchanged.

RPC limitation: direct steer and follow_up commands bypass Pi's input hooks, so their shorthand stays literal. Send prompt with streamingBehavior: "steer" or "followUp" while streaming to get conversion. This extension uses public hooks rather than patching Pi's message pipeline.

Troubleshooting

• No popup: confirm the extension loaded, use an interactive terminal, and type $ at a prose token boundary. Check that the expected Skill is in Pi's discovered resources; reload after changing resource configuration.
• No link: check the exact case-sensitive name, its discovered document path, and whether the reference is escaped, in code, or already inside a link.
• Wrong target: inspect Pi's resource origins and duplicate-name precedence. The selector uses the winning catalogue entry rather than scanning separately.
• The model did not read the Skill: inspect the submitted user message first. A correct link verifies conversion; reading the document is the model's choice.