Automate me

Turn the user's recurring working preferences into one concise -mode skill. The output records how agents should work for this user. It does not automate contact with other people.

Subagents may communicate with each other, but no agent may communicate with a person. Repeat that sentence verbatim in every child prompt. No agent may post, send, reply, or comment through external chat, email, ticket, review, or social tools. If communication is useful, prepare a local draft for the user.

1. Find the existing mode

Inspect repository instructions, the active skill catalog, and documented project and user skill roots for *-mode/SKILL.md. Do not assume one host's directory layout. If a matching mode exists, update it unless the user explicitly asks to start again.

For an update, preserve rules that new evidence does not contradict. Mine only history since the file's last meaningful edit when that boundary is available.

2. Mine workspace-scoped evidence

Use only the active workspace's conversation history exposed by the host. Never scan unrelated projects or a global transcript store. If no transcript path or conversation-history capability exists, work from the current conversation and ask the user for missing preferences.

When enough history exists, split it into a few time slices and ask parallel read-only subagents to find recurring patterns. Require evidence pointers. Promote patterns repeated across slices. Treat a single occurrence as weak evidence. When engineering-mode is installed, match each subagent's reasoning profile to the job by its Delegation rule; otherwise use the configured profile.

Look for response style, autonomy limits, delegation preferences, verification standards, code and prose rules, local process conventions, and skill-authoring habits.