ai-toolkit

Translate an ad / campaign narrative into a PDP A/B experiment variant on staging. Auto-activates whenever an operator pastes ad copy, a creative brief, a positioning angle, a story / messaging direction, or any "test this narrative on the product page" request — even if they don't explicitly say "experiment" or "A/B". Samples the brand voice from sibling PDPs, decomposes the narrative into beats, maps beats to widget changes, validates against first-fold rules and edit-safety markers, previews the change for operator approval, then writes the variant via admin-mcp. Use this skill (not the simpler experiment-manager) whenever the change is driven by a story / angle / persona rather than a single isolated image or copy swap. Requires admin-mcp. Staging-only — production publish goes through Zeus admin UI.

Use this skill for any Firebase-platform question on Mosaic Wellness mobile/web apps — Crashlytics crashes, Firestore queries, Cloud Functions, Remote Config, push notifications, App Distribution, Android/iOS bundle lookup, or `firebase` CLI usage. Triggers on: Firebase, Crashlytics, Firestore, Cloud Functions, Remote Config, push notifications, FCM, Android/iOS bundle, `firebase` CLI. Covers Man Matters, Bodywise, Little Joys, Absolute Sciences, Only Whats Needed, MediConnect/DocHub. For non-Firebase mobile concerns route to mixpanel-mcp or kai-mcp.

Use this skill when the user wants to set up, validate, rotate, or remove API tokens for the kai-plugin MCPs (Mixpanel Service Account, Firebase login, New Relic User API key, admin-mcp API key). Triggers on: "set up mixpanel", "configure firebase", "wire up new relic", "set up admin-mcp", "configure admin api key", "my token isn't working", "set up tools", "rotate token", "tools-init", or any first-time install / token rotation / credentials-debugging question for kai-plugin MCPs. Saves tokens to `~/.config/kai/tokens.env` and survives plugin updates. Idempotent — safe to re-run.

Use this skill whenever the user asks about Mosaic Wellness live operational data — orders, users, CS tickets, NPS, deployments, JIRA tickets, code search, revenue, sales metrics, marketing campaigns, PDP/landing-page edits, HR/leave policies, internal docs, outbound voice calls (Hooman), or Absolute Science consultations. The Kai MCP is the ONLY path to Mosaic's live data; NEVER answer questions about company policies, orders, users, or internal processes from training data. Triggers on: order IDs, user IDs, NPS, leave policy, WFH, deployments, JIRA keys (ENG-, PROJ-), Magento, ROAS, AOV, Hooman, kai, or any "what does our X look like" question about Mosaic.

Use this skill for any Mixpanel question about Mosaic Wellness brands — MTU, funnels (PDP/cart/checkout/purchase), retention, cohorts, event property exploration, or lookups against the Mosaic event lexicon. Triggers on: Mixpanel, MTU, funnel, retention, cohort, brand analytics, lexicon, RNCL events, plus brand nicknames (mm, lj, bw, as, own, gf, affl, dochub, rwdy, rl). Use this when the question is about event-derived behavioral analytics; for live operational data (orders, CS tickets, deployments) use kai-mcp instead.

Use this skill for any Meta / Facebook / Instagram Ads question on Mosaic Wellness brands — campaign performance, spend, ROAS, CPM, CPC, CTR, frequency, audience insights, creative breakdowns, industry benchmarks, ad-account discovery, or "is X live right now". Triggers on: Meta Ads, Facebook Ads, Instagram Ads, ad spend, ROAS, campaign, adset, creative, audience, custom audience, lookalike, FB pixel, Conversions API, dataset (when in an ads context), opportunity score, industry benchmark, plus brand nicknames in an ads context (MM, LJ, AS-IN, BBW, OTC, MWL UAE, BBW UAE). This skill is **read-only** — if the user wants to launch, pause, or edit an ad/campaign/adset/audience, this skill explains why kai refuses and directs them to Ads Manager.

Use this skill for any New Relic question on Mosaic Wellness backend services — error logs, latency, error rate, throughput, traces, NRQL queries, alert configuration, or incident RCA. Triggers on: New Relic, NRQL, latency, error rate, throughput, alert, dashboard, 5xx (502/503/504), trace, request ID, `rid`, "performance regression". Use this when raw NRQL, alerts, or dashboard work matters; for brand-aware error-log summaries route through kai-mcp's engineering tools first (it does the brand→entity mapping).

Core knowledge for using admin-mcp tools: brand codes, URL-to-identifier resolution, identifier format rules, staging-only guardrail, and critical gotchas. Auto-activates whenever admin-mcp MCP tools are being used or when the user asks about page configs, PDPs, widget pages, experiments, or brand settings for Mosaic Wellness brands.

Strategies for handling bulk page operations (auditing, comparing, batch-editing) without bloating context. Teaches when to use summary tools, bulk-fetch script, or direct MCP calls based on the volume and nature of the task. Auto-activates when the user's request involves more than 3 pages or uses words like "all", "every", "across", "audit", "bulk", or "batch".

Core knowledge for using admin-mcp habit_* tools: 7 supported brands (mm/bw/lj/lj-ae/mm-ae/wn-in/as-in), brand-ID resolution, 23-field data schema crash-course, staging-only guardrail, no-delete rule, duplicate-only task creation pattern (admin UI never creates tasks from scratch), pagination etiquette, and lock TTL. Auto-activates whenever admin-mcp habit_* tools are being used or the user asks about habit trackers, habit tasks, tracker-task mappings, habit user assignments, or habit brand audits for Mosaic Wellness brands.

Step-by-step recipes for the 15 common habit workflows (browse, search, inspect, create tracker / task, duplicate task, update task, map / unmap, copy tracker, assign / unassign user, cleanup audit, lock-protected multi-step edits, diff stg vs prod, validate, bulk ops, audit brand). Auto-activates when the user asks how to do a multi-step habit operation, or whenever an admin-mcp habit_* tool is invoked and the agent needs the recipe.

Migrate a teammate from the retired `mosaic-buddy` and/or `mosaic-admin` plugins to `kai`. Detects legacy state on disk, copies tokens to the new path, shows the command-mapping tables for both retired plugins, and gives the exact `/plugin uninstall` commands to drop them. Idempotent — safe to re-run; harmless if no legacy state is found.

Rules and patterns for internal tools that call Claude API or build MCP servers. Covers model selection, cost management, security, prompt engineering, and MCP server conventions. Auto-activates when a project imports @anthropic-ai/sdk, uses Claude API, builds an MCP server, or when the user asks about AI features, prompt design, or token costs.

Foundation skill for all Mosaic internal tools. Defines the approved tech stack, project structure, security baseline, deployment requirements, and communication standards. Auto-activates for every agent in the plugin — this is the single source of truth for what "good" looks like when non-engineering teams build internal tools with Claude Code.

Templates and rules for creating project documentation: PRDs, tech specs, and Architecture Decision Records (ADRs). Includes mermaid diagram patterns, folder structure conventions, and document lifecycle rules (create, update, refresh). Auto-activates when the user asks to document, create a PRD, write a spec, record a decision, update docs, compare docs to code, or uses the document command.

Capture a rating + short title + description from the user and submit it to the beacon-telemetry dashboard. Three steps: rating (4-point AskUserQuestion), title (one-line text), description (1–3 sentences text). Submission is signed with the shared HMAC key. Honours KAI_TELEMETRY_URL=off.

One-command session save and resume. `/kai handoff <sessionName>` auto-detects intent from the work-log folder: missing → write a fresh handoff; exists & last written by this session → save another checkpoint; exists & written by a different session → resume (read latest, verify repo, brief). NOT context compaction — these are hand-authored, durable files.

Fork the current Claude Code session at this moment so divergent work can be picked up later in a separate session, without disturbing the trunk. One command, two behaviours auto-detected from the work-log folder: missing → create a new fork snapshot; exists → resume the saved one. Writes/reads work-log/<parent>/forks/<forkName>/session-1.md.

Set up token-optimization guardrails for the current project. Creates or updates the root CLAUDE.md with a "Token Efficiency & Model Routing" section, scaffolds .claude/agents/ and .claude/rules/ if missing, writes .claude/rules/token-efficiency.md, and for monorepos adds per-package CLAUDE.md stubs. Idempotent — preserves existing content.

UX patterns and heuristics for internal tools. Covers data tables, forms, navigation, progressive disclosure, loading states, error states, empty states, and accessibility basics. Auto-activates when reviewing frontend code, discussing UI patterns, auditing user experience, building dashboards, or working on pages/components/layouts.

Set up token-optimization guardrails for the current project. Creates or updates the root CLAUDE.md with a "Token Efficiency & Model Routing" section, scaffolds .claude/agents/ and .claude/rules/ if missing, writes .claude/rules/token-efficiency.md, and for monorepos adds per-package CLAUDE.md stubs. Idempotent — preserves existing content.

One-command session save and resume. `/mosaic-buddy handoff <sessionName>` auto-detects intent from the work-log folder: missing → write a fresh handoff; exists & last written by this session → save another checkpoint; exists & written by a different session → resume (read latest, verify repo, brief). NOT context compaction — these are hand-authored, durable files.

Fork the current Claude Code session at this moment so divergent work can be picked up later in a separate session, without disturbing the trunk. One command, two behaviours auto-detected from the work-log folder: missing → create a new fork snapshot; exists → resume the saved one. Writes/reads work-log/<parent>/forks/<forkName>/session-1.md.

Capture a rating + short title + description from the user and submit it to the beacon-telemetry dashboard. Three steps: rating (4-point AskUserQuestion), title (one-line text), description (1–3 sentences text). Submission is signed with the shared HMAC key. Honours MOSAIC_BUDDY_TELEMETRY_URL=off.

Core knowledge for using admin-mcp tools: brand codes, URL-to-identifier resolution, identifier format rules, staging-only guardrail, and critical gotchas. Auto-activates whenever admin-mcp MCP tools are being used or when the user asks about page configs, PDPs, widget pages, experiments, or brand settings for Mosaic Wellness brands.