ideate
The front door for fuzzy ideas: an open design dialogue — context first, decomposition before detail, 2–3 approaches with a recommendation — ending in a user-approved design handed to define-goal. Never writes goals or code.
Four skills-first plugins turn requests into measurable goal contracts, rich HTML artifacts, autonomous metric optimization, cleaner human-sounding prose, and one-goal-at-a-time commits kept only after local checks pass. The workflow is the “proactive loops” pattern from Anthropic’s official loops guidance, shipped as installable skills.
The marketplace pairs flywheel's factory chain with three focused sidecar plugins — html-artifacts, autoresearch, and human-writing. flywheel explores fuzzy ideas into designs, captures wants as contracts, queues them in the repo, works ready goals one at a time (one per run by default; batch flags available), and keeps each commit only after TDD-backed local checks pass.
A fuzzy idea → a user-approved design through open dialogue. Hands the design to define-goal; never writes goals or code.
Plain-language want → a measurable contract, grounded by recon and red-teamed before it queues into docs/goals/.
Goal files + index.yaml live in the repo. Status lives only in the index, so a goal is never worked twice.
Works ready goals one at a time on the current branch — no worktrees, no PRs. One foreground implementer, backed by TDD and fresh review — then the orchestrator re-reviews the diff with its own independent reviewer.
Each goal's commit is kept only after an independent review and a local build+test gate pass; a failure rolls it back. CI is a non-blocking post-push check.
Install flywheel for the goal queue and autonomous execution. Install html-artifacts for browser-file deliverables, autoresearch for autonomous metric optimization, and human-writing to strip AI-writing tells. In Claude Code they're namespaced by plugin.
The front door for fuzzy ideas: an open design dialogue — context first, decomposition before detail, 2–3 approaches with a recommendation — ending in a user-approved design handed to define-goal. Never writes goals or code.
Turns a plain-language want — or a whole document of them — into a measurable goal contract, grounded by recon (Sonnet gather agents, session-model synthesis) and a concise brief. Red-teams every queued contract before it lands, then stamps the implementer model last — opus by default for features and bugs, sonnet for mechanical work. Produces goals only; never implements.
Separate plugin for self-contained browser artifacts: rich plans, specs, PR reviews, diagrams, research explainers, decks, prototypes, and one-off editors with export/copy round trips.
The factory orchestrator. Works ready goals one at a time — the next goal, an exact goal (/dispatch 087), or a sequential batch (--count 3, --unlimited) — commits straight to your branch, and keeps only what passes an independent second-view review plus the local build+test gate. Orchestrates only.
A read-only glance at the queue: every in-progress, blocked, and not-started goal with its title and brief, grouped in that order — completed goals hidden. Blocked goals show their reason; a goal waiting on a dependency shows what it waits on. Never changes the queue.
Designs the loop contract — prompt + verification + stop conditions — asking only the calibration needed so an unattended run converges and stops. Includes usage-limit proofing: scheduling that survives the account’s 5-hour/weekly limit windows.
One-pass preflight for a repo + machine: software, a clean working tree, the working branch, the verify commands, queue state, and loop health — stale claims, underspecified goals, usage-limit exposure. Auto-fixes everything local.
Separate plugin for an autonomous optimization loop: try a hypothesis, measure it, keep improvements and revert regressions, with MAD-based confidence scoring and file-based logs any session can resume.
Separate plugin that edits AI-sounding text into human prose: strips inflated significance, promotional language, em-dash and rule-of-three overuse, AI vocabulary, and chatbot artifacts. Based on Wikipedia's "Signs of AI writing".
No issue-body size limits, no per-repo label bootstrap — versioned with the code. The branch you're on is the integration surface.
# status lives ONLY in index.yaml — goal files are immutable contracts docs/goals/ ├── index.yaml # config + queue state ├── 001-faster-checkout.md ├── 002-fix-auth-redirect.md └── done/ # archived completed goals
config: base: staging # the branch goals are committed to model: inherit # code-agent default — a goal's own model: wins skills: [test-driven-development] # repo-wide mandates verify: # local build+test gate (ordered) - npm run build - npm test budget: # optional external brake max_iterations: 20
Decided from real production runs, encoded so an unattended loop advances the queue without corrupting it.
single sourceStatus lives only in index.yaml; dual-writing frontmatter drifts.claim protocolFlip one entry to in_progress before working it, so a goal is never picked up twice.direct-to-branchThe implementer commits straight to the checked-out branch — no worktrees, no goal/<id> branches.blocked ≠ lostA blocked goal keeps a reason and is surfaced under needs-you, never re-dispatched into livelock.local gateOne goal per run; the commit is kept only after TDD-backed verification and the verify build+test gate pass, else rolled back.pragmatic-growth marketplace.One marketplace, four plugins.
/plugin marketplace add pragmaticgrowth/flywheel /plugin install flywheel@pragmatic-growth /plugin install html-artifacts@pragmatic-growth /plugin install autoresearch@pragmatic-growth /plugin install human-writing@pragmatic-growth
/plugin marketplace update pragmatic-growth
/factory-doctor # preflight the repo + machine /define-goal I want the API p95 latency under 200ms /dispatch # work one ready goal /goals-status # see what's still open