Skill Autopilot v1.0.0

Skill Autopilot — use the skills you've installed, not just the ones you remember.

Formerly Command Autopilot. v1.0.0 renames the project and repositions it around your installed skills (the old repo URL redirects automatically).

What it does

Each turn, the plugin ranks your installed skills against what you're actually asking and surfaces the few most relevant names — the model reads each full description on demand and decides whether to use one. This reaches skills Claude Code dropped from context for being over its skill budget: the perfect-but-forgotten skill that would otherwise be invisible to the model. It also lays out the high-leverage moves that fit a big task as a menu you pick from, and handles the basics (undo, context hygiene) quietly.

Honest framing

New and experimental. The relevance ordering is attention-narrowing, not a verdict the model is told to trust — matching stays the model's job. Effectiveness is not yet proven on real-world data; the local ledger is how it will be measured and tuned. No effectiveness numbers are claimed. Zero telemetry; everything is local files you can read, audit, and delete.

Install

claude plugin marketplace add WinterDDo/claude-code-skill-autopilot
claude plugin install skill-autopilot@claude-code-skill-autopilot

Then fully restart Claude Code and ask: "check that the autopilot is working."

49 unit + 3 smoke tests; claude plugin validate passes. MIT licensed, local-first.

v0.6.0 — a menu of high-leverage options, not a single pick

At a genuine high-value moment the better-move check now presents the 2-4 fitting high-leverage moves as a clickable menu (command + payoff) + 'no, continue', so you see and choose your options. Same silence-default gate (not more frequent, just more useful when it speaks). Drops the redundant per-command /goal hint (the general rule already recommends the right command ~88% across the surface).

v0.5.1 — /goal cue for long autonomous-to-PR tasks

Adds 'a long task to drive autonomously to a PR (/goal)' to the better-move check's observable signals, so long 'do it all to a finished PR' tasks reliably get a /goal recommendation up front. Validated 10/12 -> 12/12 on that scenario, no new false-positives. One phrase, no new machinery.

v0.5.0 — BETTER-MOVE CHECK (first-principles redesign)

First-principles refocus: the co-pilot IS the per-turn prompt, not a pile of triggers. The whole value is a few well-calibrated words that make a capable model surface a better move when (and only when) one genuinely helps — for anyone, beginner to expert.

What changed

• The core rule is now a BETTER-MOVE CHECK. Every turn, before committing to an approach, the model silently asks whether a materially better move exists for THIS task — a command, combo, skill, workflow, or a pattern from your own past. Default is silence; it surfaces at most one move, and only when it clears a high bar (concrete · not-already-chosen · materially better · non-obvious now), biased to observable moments: a risky/irreversible op, a fitting capability you've never used, visible repetition, being stuck. Cheap + reversible moves it just takes; costly / control-handover ones it RECOMMENDS as a clickable choice and lets you decide.
• Removed the v0.4.x PostToolUse background "moment-nudge" hook (and its cloud sh). A full day of real use showed it fired ~0 organically while adding complexity; the every-turn rule covers the same moment earlier and more reliably. Net change: the codebase got simpler (−276 lines).

Why it matters for everyone, not just beginners

Forgetting your own best practice — not ignorance of a command — is the expert's real failure. So recalling your own pattern at the moment it applies is a first-class move. Honest scope: it does not claim to pick the optimal move every turn; it is a mostly-silent co-pilot that surfaces rarely and precisely, and gets quieter as it learns you.

Install

claude plugin marketplace add WinterDDo/claude-code-command-autopilot
claude plugin install command-autopilot@claude-code-command-autopilot

Then fully quit and reopen Claude Code (hooks load at startup).

Full changelog

CHANGELOG.md

v0.4.1 — cloud copilot present every turn

The cloud copilot is now present every turn, not just once — and a real cloud session just confirmed it firing per prompt.

What's new in 0.4.1

The vendored cloud hook used to inject the autopilot rules only at SessionStart. In a long session that aged out, so cloud never surfaced the high-leverage moves (Workflow fan-out, /goal, /loop, parallel /fork, background tasks). It now wires both events to one script:

• UserPromptSubmit — per-prompt, exactly matching the local plugin.
• SessionStart — a guaranteed-to-run fallback.

The script stamps which event fired into its own output. Worst case is the old behavior; best case is full per-prompt parity with local. The guard now keys on $CLAUDE_CODE_REMOTE: cloud always injects, while locally with the plugin installed it stays silent so you never get a double.

Confirmed working in a real cloud session. The cloud runs the repo-committed UserPromptSubmit hook, giving full per-prompt parity with the local plugin. This is verified, not theoretical.

Why per-turn presence matters

You can't ask for what you don't know exists. The whole point of the autopilot is to surface the powerful move at the moment it would help — before you grind through a big, repetitive, or long-running job the slow way. A copilot that speaks once at session start and then goes quiet misses every one of those moments later in the session. Per-turn presence means the offer shows up when the task actually calls for it, in the cloud and for teams, not just in the first message.

Install

claude plugin marketplace add WinterDDo/claude-code-command-autopilot
claude plugin install command-autopilot@claude-code-command-autopilot

Then fully quit and reopen Claude Code (hooks load at startup), and ask Claude "check that the autopilot is working".

Full changelog

CHANGELOG.md — entries 0.1.0 through 0.4.1.