ramble

ramble — run a human-feedback loop through RambleDesk

Route anything that needs a real human's judgment, review, approval, or hands-on experience through the RambleDesk feedback loop instead of asking in chat.

Invocation contract

• Treat explicit invocation as a command, not a hint. If the user invokes /ramble, $ramble, [$ramble], names this skill as the way to handle the task, or asks for RambleDesk feedback, create the first RambleDesk request before substantive advice, design, review, or implementation.
• Treat /ramble [task] as a task-scoped loop. Guarantee the first request, then create later serialized requests only when the same task needs another clarification, review, or final confirmation. Do not carry the loop into an unrelated future task.
• Treat an empty argument or a generic starter such as "start this Ramble" or "Let's work on something together" as no meaningful task:
  • If the conversation already contains one clear active task, summarize the current understanding in the kickoff request and ask the human to confirm or correct it while supplying any missing context, constraints, desired output, priorities, and completion criteria.
  • Otherwise create a kickoff request asking the human for the goal, relevant context and materials, constraints, desired output, priorities, and completion criteria.
  • Ask for this task brief in RambleDesk, never in the host chat.
• Treat /ramble_on and /ramble_off, when the host provides them, as persistent-mode switches rather than one-task skill invocations.
• Use this skill implicitly when human testing, visual inspection, screenshots, clarification, or a decision would materially improve the result.
• Do not treat a meta request to inspect, edit, or debug this skill as a request to create feedback unless the user also explicitly asks to run the loop.

Select the available flow

If the current session includes RambleDesk managed workflow context (or RAMBLEDESK_MANAGED_SESSION=1), follow that context first. Use the application command supplied by RAMBLEDESK_COMMAND: feedback request --input <json-file>, then end the Agent turn. RambleDesk resumes this same conversation after human feedback; use feedback get --request-id <original-id> to read it. Use feedback recover after an interruption. Do not select the native adapter or Generic MCP flows below, do not poll or block, and never choose host/session identifiers yourself. The private capability is supplied by the runtime; missing credentials require reconnecting this session, never external fallback.