Requesting new skills (conversational creation)

When no existing skill covers a need, your agent asks the platform to author one. Creation quality is bounded by brief quality, so the platform is deliberately conversational about it.

The golden rule: terse on reads, conversational on writes

Searches are cheap one-liners. Creation briefs are the opposite — write to the platform the way you'd brief a skilled colleague:

• the task narrative — what you're doing and why the capability is needed;
• what you already tried, and why near-matches fell short;
• one or two concrete input/output examples from the real task;
• environment, formats, edge cases, and success criteria.

Bad:   skill_request(need="parse invoices")

Good:  skill_request(
  need="parse supplier invoices into a ledger schema",
  context="Working through ~200 PDF invoices from 12 suppliers in
    data/invoices/. Need line items with VAT split out; output must match
    ledger/schema.json. Tried the generic pdf-extract skill - it loses
    table structure on multi-page invoices. Success = every invoice in the
    sample set round-trips with totals matching. Input example: ...
    Output example: ...")

Clarifying questions

If the brief is too thin, the platform asks before it writes:

• Clients with MCP elicitation (Cursor, Claude Code and others): questions arrive mid-request; your agent answers from its real task context and the same request continues to the result.
• Clients without elicitation: the response comes back as action: "clarify" with the questions. Answer them all in one string and re-call with the same need plus your answers in the answers parameter.

Reviewing the draft before it publishes

For high-stakes skills, pass review: true (elicitation clients only). The platform shows the draft before anything is persisted; you approve it, request one round of changes, or reject it outright.

Outcomes