Verified 2026-09-23
view as .md
Runtime · 1 rule

How should my skill ask the user for missing arguments?

When your skill is invoked, the host injects an instruction telling the model to collect anything it still needs through an inline elicitation form rather than through AskUserQuestion, and the answers come back as bullet points in the next message instead of as a tool result. Declaring an argument-hint is what makes that path work well. This is a steer, not a wall: AskUserQuestion still exists and still functions.

Do not build your skill's argument collection around AskUserQuestion in Cowork.

FRICTIONIt works, but costs the user an extra prompt or step.From binaryRead from a shipped artifact (the app, the agent or the sandbox image). The behaviour itself was not exercised. Can change without a version bumpServed from Anthropic's side, so it can change without an app update.

On every top-level skill invocation (not inside sub-agents) the host injects an instruction telling the model to gather missing context through the elicitation form — pills, free text, dates, a file dropzone — and explicitly not through AskUserQuestion, which it reserves for one-off clarifications mid-task. The answers arrive as bullet points in the model's next message, not as a tool result, so a skill written to wait for a tool result is waiting for something that will not come. Declaring argument-hint in your frontmatter is what makes this path work: the hint is passed through verbatim, and a skill without one leaves the model told to infer what to collect from your SKILL.md.

Caveat: This is a steer, not an enforcement boundary. AskUserQuestion remains callable and functional; what was verified is that Cowork injects the instruction, not that a model obeys it in any particular run.
Caveat: Delivery is gated server-side and read on at the 2026-08-13 and 2026-09-23 captures. A session where that flag is off receives no such instruction, and the guidance does not apply.
Caveat: Measured on one machine: in 71 Cowork tasks that asked the user anything after receiving the instruction, the model used the form in 31, AskUserQuestion in 27 and both in 13 — so your skill must handle answers arriving either way.

Checked live 2026-09-23 · Desktop 2.7032.0 · agent 2.1.280

What is not established

  • The instruction is injected text, not an enforcement boundary, and the model usually but not always follows it. On Desktop 2.19675.0, in every founder-skills session on Opus 5.5 that asked anything, the first question went through the form in 13 of 14, and in both such sessions on Opus 4.8. Other skills, models and builds were not counted.