Design and analyze A/B tests for AI agents — compare prompt versions, model choices, or architecture variants with statistical rigor. Use when optimizing agent performance, testing new instructions, evaluating model upgrades, or validating TK-based improvements.

Create demo videos for AI agents using Remotion (React-based video framework). Compose screen recordings, architecture animations, narration, and subtitles into polished demo videos. Use when showcasing agent capabilities to stakeholders, investors, or users.

Design a Go-To-Market strategy for AI agent products — identify beachhead segments, define launch sequence, and plan adoption motions. Use when preparing to launch an agent product, choosing first customers, or planning the transition from internal tool to external SaaS.

Review agent plans (PRD, architecture, instructions) before implementation — 4-axis verification: scope check, architecture, instruction quality, operational reliability. Use when validating an agent design before committing to build, reviewing PRDs, or checking architecture decisions.

Identify and prioritize the riskiest assumptions in an agent idea across four axes: Value, Feasibility, Reliability, and Ethics. Use after defining an agent opportunity and before starting implementation. Prevents building agents that work technically but fail operationally or cause unintended harm.

Design a sustainable business model for AI agent products — pricing strategy (per-use, subscription, outcome-based), cost structure, unit economics, and value capture. Use when planning agent monetization, evaluating pricing models, or building a business case for agent development.

Decide whether to build a custom agent, buy/subscribe to an existing solution, or use a no-code/low-code platform. Use before committing to agent development — this decision affects cost, time-to-value, differentiation, and long-term maintenance burden. Prevents the 'build everything from scratch' default trap.

Track, analyze, and optimize token costs for AI agent operations. Break down costs by model, task type, and user segment. Use when reviewing monthly agent spending, planning cost reduction, comparing model economics, or setting token budgets for new agents.

Scan a project's structure, tech stack, and conventions — then generate a production-ready CLAUDE.md that turns Claude Code into a project-aware teammate. Also recommends the right hplan plugins and skills based on what the project actually needs. Use as the first step when onboarding Claude Code to any project, or when an existing CLAUDE.md feels incomplete.

Executable COGS gate for AI products. Runs a deterministic Python sampler (lognormal token-cost distribution) to compute p50/p90 per-paid-user monthly COGS, gross margin scenarios, and free-user abuse blend. Returns GREEN / CONDITIONAL_GO / RED before any paid AI product is greenlit. Use when promising a paid AI feature, comparing providers (Anthropic/OpenAI/Google), or when oracle/cost-sim has produced a usage hypothesis and you need real numbers.

Simulate and forecast agent operating costs before building. Model token consumption, API call frequency, and monthly burn rate across different models and usage patterns. Use when evaluating agent feasibility, setting cost KPIs, or comparing build vs buy economics. Prevents the 'it's just API calls' cost surprise.

Decide which agent should handle an incoming request across multiple teams — based on tier, capability map, current load, and inter-team handoff cost. Use when a request could plausibly be served by 2+ agents from different teams, when load-balancing across an agent fleet, or when introducing a new agent and deciding which team's traffic it absorbs.

Plan and optimize the context window usage for an agent. Define file loading priority, estimate token consumption per component, and design a strategy to stay within budget. Use when designing agents that load multiple files, or when diagnosing high token costs or context overflow errors.

Decompose a PRD or feature into a WBS, classify each task complexity 1-5 (LLM classification, Rule 5 허용 영역), then resolve (expected_loc, tokens_p50/p90, minutes_p50/p90) by deterministic percentile lookup from profiles velocity baseline — NOT by LLM token hallucination. Outputs .track/predicted.json that locks the baseline for progress-probe to measure deviation against. Use at the start of any track-init flow or when scoping a new feature for parallel-team.

Score a product idea against the 100-point evidence rubric before any PRD work. Eight axes: ICP specificity, recent painful event, current workaround, repetition, economic pain, switching trigger, MVP narrowness, and acquisition path to first 5 users. Returns build/interview/pivot/hold decision plus the specific axes that are weak. Use when a founder or PM is excited about an idea but evidence is thin, or before approving any spec-driven coding workflow (Spec-Kit, Kiro, GStack, Superpowers).

Design the build harness for an agent project — a 4+ agent team, an autonomous Ralph Loop, dry-run + backup gates, and pending-input batching. Use when scope crosses a single PR's worth of work, when implementation requires iterative self-correction, or when the user has approved 'autonomous mode' (ralph loop) with question-free batching. Strictly distinct from architect/orchestration which designs the runtime system, not the build process.

Measure visual hierarchy rules at runtime via Playwright + DOM saliency + WCAG AA — fold_density (count of elements above 1080px viewport fold), type_hierarchy (unique font sizes + consecutive size ratios ≥ 1.25), color_60_30_10 (pixel-level color distribution clustered to 3 primary colors), whitespace_ratio (empty pixels / content pixels ≥ RESPECT.md threshold), cta_count_above_fold (exactly 1 per RESPECT.md). All thresholds are loaded from RESPECT.md — this skill is the runtime enforcement arm. Doubles as data supplier for track/respect-checkpoint γ gate.

Create animated infographics (GIF/MP4) from HTML/CSS to visualize agent architectures, workflows, and data flows. Use when you need visual explanations for stakeholder communication, documentation, or demo materials. Routes to compose-video for multi-scene final rendering.

Define and track Key Performance Indicators for AI agents — operational metrics (latency, success rate, error rate) and business metrics (task completion, user satisfaction, cost per task). Use when setting up agent monitoring dashboards, defining SLAs, or establishing performance baselines.

Design an agent memory system — working memory, episodic memory, semantic memory, and procedural memory. Use when building agents that need to remember context across sessions, learn from interactions, or maintain persistent knowledge. Covers storage strategy, retrieval patterns, and context management.