Autoselect Skill

Automatically recommends and ranks relevant skills to load based on the user message's implied domain or workflow, with an opt-in trigger.

Purpose

When a user message arrives, this skill scans the set of available skills and recommends which ones to load for the current turn. It does NOT replace loading those skills — the agent then calls skill for each recommended name.

When to invoke

Auto-trigger (heuristic) — fire this skill when the user message:

• Names a domain with high specificity (e.g. "review my UI/UX", "research market trends", "debug this stack trace", "sell into X company")
• Implies a workflow class (planning, design, deep research, code change, doc generation, image/visual)
• Mentions a tool name that overlaps a skill (e.g. "Mermaid", "Vite", "Feishu")

Manual trigger — user types /autoselect. Always run.

Skip — if the message is pure small-talk, single-question, or already targets one skill explicitly (/plan-mode, /skill foo).

Workflow for the invoking agent

1. Collect the list of available skills. The runtime injects this at session start as <available_skills> with each name and description. If not injected, call mavis({ command: "agent list" }) and inspect prompt frontmatter.

2. Score each skill 0–3 against the user message:

   Score	Signal
   3	Skill name appears in the message (verbatim or stemmed)
   2	Skill description keyword overlap ≥ 2 of the message's content words
   1	Adjacent domain — could plausibly help but not primary
   0	No overlap — exclude

3. Cap at top 3 skills with score ≥ 1. Drop the rest.

4. Output format (always emit this block when called):

   text

   <autoselect>
   recommended_skills:
     - name: <skill-name>
       score: <0-3>
       reason: <one sentence tying skill description to the user message>
     - name: <skill-name-2>
       score: <0-3>
       reason: <one sentence>
   ignored_skills:
     - name: <skill-name>
       reason: <why skipped, e.g. "score 0, no overlap">
   next_action: load <skill-name>[, <skill-name-2>...] OR proceed without loading if all scores are 0
   </autoselect>
   

5. After emitting the block, the agent may then call skill for each top recommendation to actually load its body into context.

Worked example

User: "Can you review this UI mockup and propose a better color palette?"

text

<autoselect>
recommended_skills:
  - name: ui-ux-pro-max
    score: 3
    reason: "UI + palette + design review explicitly named in the message"
  - name: visual-content-generator
    score: 1
    reason: "Adjacent — could generate alternative mockups, but not strictly asked"
ignored_skills:
  - name: deep-research-agent
    reason: "score 0, no research ask"
next_action: load ui-ux-pro-max
</autoselect>

Universal — works for any LLM

This skill does NOT depend on Mavis-specific runtime features beyond the <available_skills> injection. Any LLM with:

• A way to list its skills
• A way to load a skill (file / system prompt / tool call)

can run this workflow by following the steps above.

Cross-project use

• Treat the output as advisory. The invoking LLM retains final judgment — if its own reading of the message diverges from the heuristic, trust the reading.
• Prefer precision over recall: an empty <autoselect> block (no recommendations) is a valid answer. Over-triggering adds noise.
• Update the trigger list above if new domain-bridging skills are added to the roster.

Install

Drop SKILL.md into your agent's skills directory. For Mavis-format agents, that is $AGENT_HOME/.skills/autoselect-skill/SKILL.md. Restart the agent so the new skill is picked up by the skill loader.

License

MIT — same as the parent project.