Proactive — a Claude Code output style (+ skill)

Make Claude self-drive multi-step work to completion and surface the next step, risks, and gaps before you ask — instead of stopping after every sub-task to ask "should I continue?".

If you've ever typed "continue", "keep going", or "why did you stop?" — that's the itch this scratches. Every time you have to say continue with no real decision to make, that's an orchestration failure. Proactive removes it.

What it does

• Self-drive — runs a task lane end-to-end (task → verify → next → done) without pausing between sub-steps to ask permission. A recommendation is treated as a decision, not a question.
• Anticipate & surface — names the next step, likely risks, related work, and design gaps before you have to ask.
• Safe by construction — still stops and waits at hard-gates: irreversible or outward-facing actions (settings writes, pushing to main, PRs, rm -rf, exposing secrets/PII, a strategic fork only you can decide).

It keeps normal engineering discipline (surgical changes, match existing style, verify before claiming done). Proactivity is layered on top — not a licence to be reckless.

Two faces

Form	When it runs	Use it for
Output style (recommended)	Always-on, every response	Claude proactive by default in a project/session
Skill	On-demand, per task	Flip proactive mode on for one specific task

Same behavior, two delivery mechanisms. The output style is the primary form — "proactive" only makes sense as an always-on posture — and the skill is the on-demand face for people who'd rather toggle it per task.

Install (plugin)

/plugin marketplace add peterphung/Agent-Proactive-Output
/plugin install agent-proactive@agent-proactive-output

Then:

• Always-on: /output-style proactive
• Per task: invoke the proactive skill

Install (manual, no plugin)

Just want the always-on output style:

mkdir -p ~/.claude/output-styles
curl -sL https://raw.githubusercontent.com/peterphung/Agent-Proactive-Output/main/output-styles/proactive.md \
  -o ~/.claude/output-styles/proactive.md

Then in Claude Code: /output-style proactive.

Use it in other tools (CLAUDE.md · Cursor)

The same behavior ships in portable, tool-agnostic forms — no plugin needed:

• Any tool with a root instruction file: copy CLAUDE.md into your project (or merge it into your existing instructions).
• Cursor: copy .cursor/rules/proactive.mdc into your project's .cursor/rules/ — it's committed with alwaysApply: true. See CURSOR.md for details.

The output style is canonical; CLAUDE.md, the skill, and the Cursor rule are copies of the same instructions for different tools.

What it will still stop for (hard-gates)

Proactive self-drives reversible, reported work. It always pauses for:

• writing settings/config files you didn't direct
• pushing to a shared branch (main) · creating PRs · merging its own PR
• irreversible deletes (rm -rf) · deleting skills
• exposing secrets or PII
• a genuinely strategic fork only you can decide

Examples

See EXAMPLES.md for before/after transcripts.

License

MIT — see LICENSE.