personal-mode

Personal Mode

Project skills say how to work in this repo. A personal mode says how you work, everywhere: the calls you would make, the depth you want, the ceremony you skip. This skill drafts one from evidence — your recent transcripts — not from a questionnaire.

When to use

• the user asks for agents to work like them across projects, or to remember their working style
• the same correction keeps coming up in unrelated repos

When not to use

• repo-specific conventions — those are project skills (skillify)
• one-off preferences — remember stores those cheaply

Workflow

1. Gather evidence: recent transcripts across the user's projects (with the user's say-so for which span). Look for repeated signals: corrections the user made more than once, reply-length or verbosity preferences, risk calls (autonomy grants, pause points), tools they always or never want used.
2. Draft the profile as skill edits, grouped and each tied to the transcript moments that evidence it. A preference seen once is a hypothesis; state it as one and ask.
3. Confirm with the user — every line of a personal mode is a standing call the user is making about their own agents; nothing lands unconfirmed.
4. Write the confirmed profile to a separate user skills entry, ${CLAUDE_CONFIG_DIR:-$HOME/.claude}/skills/personal-mode-profile/SKILL.md. Keep this generated profile separate from the bundled personal-mode workflow so the generator remains discoverable and editable. A user skill is opt-in context; it is not a global instruction applied to every task. Use frontmatter, triggers, and short checkable rules — a personal mode that restates platitudes is dead weight (minimal-prose-discipline applies).
5. Report what was captured, what was hypothesized but unconfirmed, and how to edit it later (skill command).

Rules

• Evidence first: every profile line cites the transcript moments behind it; a line without evidence is asked about, not written.
• The profile layers under project skills — a repo's explicit rules win.
• Personal modes contain no secrets, tokens, or private data — only working-style calls.
• Re-running this skill revises; it never silently overwrites — show the diff and confirm.