flywheel

Factory dispatcher — use when the user says "/dispatch", "run the factory", wants the docs/goals queue worked, or wants to work one specific queued goal in this session ("work goal 005"). Works in any repo with a docs/goals/ queue. Works ONE ready goal per run, on the currently checked-out branch — no pull requests, no worktrees, no parallel implementers. Orchestrates only — never implements in its own context; the phase procedure lives in the skill body, never in this description.

Use when the user runs "/goals-status" or asks what's in the docs/goals queue — what's left, what's blocked, what's in progress, or "show me the open goals". Prints a read-only view of the OPEN goals (in_progress, blocked, not_started), each with its title and brief; completed goals are hidden. Shows the queue only — never claims a goal, changes queue state, or implements anything (that's /dispatch).

Use when the user states something they want — a goal, wish, feature, fix, or annoyance ("I want…", "set a goal", "/define-goal"), asks to clarify success criteria or turn fuzzy intent into a measurable objective, OR hands over a document with multiple items (bug report doc, feedback list, meeting notes) to convert. Also use to add work to the docs/goals queue. Defines goals only; never starts the implementation work.

Use when the user wants to automate, schedule, or run work autonomously/remotely — setting up a /goal, /loop, routine, channel, or long unattended run, or asking "how do I keep the agent working on X". Designs the loop contract (prompt + verification + stop conditions) instead of just running the task.

Autonomous experiment loop for optimization research. Use when the user wants to: - Optimize a metric through systematic experimentation (ML training loss, test speed, bundle size, build time, etc.) - Run an automated research loop: try an idea, measure it, keep improvements, revert regressions, repeat - Set up autoresearch for any codebase with a measurable optimization target Implements the autoresearch pattern with MAD-based confidence scoring, git branch isolation, and structured experiment logging.

Use when the user asks for a non-trivial plan, spec, report, explainer, research synthesis, comparison, PR/code-review writeup, diagram, flowchart, timeline, roadmap, dashboard-like data view, slide deck, visual design exploration, prototype, playground, or one-off editor where spatial layout, color, interaction, or export back to the agent would beat markdown. Stay in markdown for short chat replies, code-only answers, command snippets, and tiny summaries.

Remove signs of AI-generated writing from text to make it sound more natural and human-written. Use when editing or reviewing any form of document including: markdown, technical docs, emails, blog posts, PRDs, or any dedicated writing content. Based on Wikipedia's comprehensive "Signs of AI writing" guide. Detects and fixes patterns including: inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary words, negative parallelisms, and excessive conjunctive phrases.

Use when setting up or troubleshooting the flywheel factory in a repo — before the first /dispatch, after a missing-queue or unauthenticated-gh error, or any time /dispatch or /define-goal behaves like the environment isn't ready. Preflights software, gh auth, CI, the local working tree, and the docs/goals queue, auto-fixing everything local. Diagnoses and fixes setup; never implements goals or merges PRs.