yangwenli

Checkpoint the finished thread, then hand back the EXACT one line to run next so the context gets cleaned and the next task starts with minimal friction. Run when one piece of work just closed and you want to move to the next task while freeing up context to save tokens. Invoke with the next task as an argument: `/rinse <next task>` (e.g. `/rinse start the Casos design day`). With no argument, rinse falls back to the **Next action** in `.claude/ACTIVE_WORK.md`. Trigger phrases: "rinse", "/rinse", "rinse and move on", "next task", "I'm done with this, what's next", "free up context and continue", "checkpoint and continue", "clean the context and start the next one", "bank this and go to <X>", "wrap up and switch to <X>". What rinse CANNOT do (be honest about this every time): it cannot clean the context or switch the model itself — `/clear`, `/compact`, and `/model` are user-only commands, and a skill runs inside the OLD context so it can't start a task in a fresh one. Rinse therefore writes the durable rec

Deploy the RUBLI codebase to production at rubli.xyz / 37.60.232.109. Use this skill whenever the user asks to "deploy", "ship to prod", "push to production", "redeploy", "deploy this fix", "make it live", or any variant. Trigger eagerly — the procedure has multiple non-obvious gotchas (VPS branch is master, not main; --no-deps frontend silently breaks the docker network; bundle hash check is the real verification) that bite every fresh session that doesn't read this first. Reading this skill before deploying saves 15 minutes of recovery the first time the site 502s.

Design-invention phase for Lylat redesign missions. DESIGNUS (Opus+ultrathink) reads the Lylat brief + live screenshots → sequential thinking → design spec → user approves inline → hands off to /starfox for implementation. Use for: M3, M4, M5, M6, M7 (full redesigns requiring design invention) Skip for: L0, M1, M1b, M2 (quick wins — invoke /starfox directly) Trigger phrases: "/mcloud", "run mcloud", "/mcloud M7", "design phase", "designus", "start lylat", "next mission", "execute mission" What /mcloud does NOT do: implement code, run gates, QA, deploy. Those belong to /starfox and /paw-patrol.

Four-agent design-modification squadron for RUBLI frontend. Unlike /paw-patrol (exploratory QA), StarFox is MISSION-DRIVEN: the user pastes screenshots showing design problems, gives instructions, and the four agents (FOX, FALCO, PEPPY, SLIPPY) analyze → implement → verify → ship. Trigger when: user pastes a screenshot AND asks for design changes, "starfox", "/starfox", "fix this design", "apply these changes", "use the squadron", "send in starfox", or any request to act on visual evidence the user provides. This skill exists because single-agent design work drifts — one agent loses track of the brief halfway through three files. Four disciplined specialists with non-overlapping domains and a commander who verifies with Playwright before shipping produce tighter, faster, and more complete design fixes.

Multi-agent exploratory QA sweep of rubli.xyz. Three independent agents (ROVER, CHASE, SKYE) navigate the live site from different entry points, screenshot evidence of issues, then converge on a prioritized fix list. Trigger when user says: "paw patrol", "/paw-patrol", "do a QA sweep", "check the site", "find issues", "exploratory QA", "browse and find bugs", "have the agents explore", or any request for unscripted browser testing. This skill exists because single-agent QA misses too much — each agent develops blind spots from their entry point. Three agents with different personas and starting locations catch orthogonal failure classes.

Audit a touched RUBLI frontend file for missing Spanish translations before claiming a redesign is done. Use this skill whenever you have edited a React/TSX file and need to verify every visible string is bilingual (`lang === 'en' ? EN : ES` pattern), when the user mentions "translate", "i18n", "Spanish", "bilingual", or whenever you are about to commit a frontend change. Bilingual gaps are the most repeated half-fix failure mode in this codebase: an agent patches the visible string and leaves two more inside event handlers, aria-labels, and tooltip builders. Running this audit before commit catches them in 30 seconds.

RUBLI's "omega" redesign workflow — the user's named multi-step pattern for redesigning specific selected UI graphics with amplified geometric changes (not annotation). Use this skill IMMEDIATELY whenever the user says "omega" (case-insensitive) about UI elements they've selected, highlighted in screenshots, or referenced via launch-selected-element blocks. Also use when the user says "redesign these specific elements", "amplify this chart", "this needs the omega treatment", or invokes similar phrasing for targeted-scope chart redesigns. The omega workflow is plan-then-execute with strict gates between steps; following it saves the cycle of "you put a little more text and now the elements are unreadable" failure mode the user explicitly named.

Bisect React error #301 ("Too many re-renders. React limits the number of renders to prevent an infinite loop.") in the RUBLI frontend. Use this skill when the prod console shows a minified React error 301, when a /atlas or dashboard page goes blank after a frontend deploy, when the user reports "the page won't load" / "it's stuck loading" / "it's blank" right after a redesign or hook change. The skill encodes the bisection procedure that recovered prod after the OMEGA-N deploy went sideways and saves the next session from re-deriving it under pressure.

Anthropic's "Three-Agent Harness" pattern (Planner / Generator / Evaluator with iterative critique) applied to RUBLI workflows. Use this skill when the user asks for a "deep" or "thorough" or "high-stakes" task, mentions "multi-agent", references the keynote pattern of "three agents cycling through plans three times", or asks to "triangulate" / "validate" / "stress-test" a design or plan. Specifically valuable for surface redesigns, risk-model retraining decisions, schema migrations, GT mining sessions, anywhere a wrong call costs hours. The pattern produces stronger plans than single-agent work because each cycle's critique catches the mid-task abandonment + decoration-not-redesign failure modes that have repeatedly cost time in this codebase.

Recovery procedure when a delegated subagent (frontend-architect, api-designer, design-visionary, etc.) abandons a task mid-flight. Use this skill whenever you spawn a Sonnet/Opus subagent for a PR-sized chunk in this codebase and the agent reports completion but the work is half-finished — missing BUILD_ID bumps, undefined symbols referenced in JSX, files written to the wrong path, or a TS error left unhandled. This pattern has bitten this codebase multiple times per session; following the recovery checklist below saves 10-15 minutes of investigation each time and avoids the failure-on-top-of-failure cascade where you rerun the agent on a mid-state worktree.

RUBLI's editorial-folio frontend aesthetic — investigative-journalism voice for a procurement corruption platform. Use this skill whenever you are redesigning, polishing, or building any user-facing surface in the RUBLI codebase (frontend/ pages, components, charts, dashboards). Trigger enthusiastically when the user says any of: "redesign this page", "make this look better", "polish this", "make it feel investigative", "add visual hierarchy", "this looks generic", "this feels like a dashboard", "make it feel like the FT/Reuters/ICIJ/OCCRP", or invokes /omega and similar redesign workflows. Also trigger when adding new pages, building charts, or working on /atlas, /dashboard, /sectors, /aria, vendor profiles, story narratives, or anything in frontend/src/pages or frontend/src/components/atlas, components/charts, components/sectors. This skill encodes the visual decisions made across many redesign cycles so future sessions ship in the right voice without rediscovering the vocabulary; reading it first saves an