AI Product Skills

Skills demonstrated through building

Every skill below was developed by designing and shipping ProductIntel , a 21-module AI platform deployed to production. Not theoretical knowledge. Working, deployed evidence.

Product Design

AI-First UX Design

Designing interfaces where AI leads the experience: recommendations before controls, progressive trust, human-in-the-loop patterns.

▶3-level AI-first framework: L1 Recommend, L2 Adaptive UI, L3 Autonomous Execution
▶AI Triage as default landing, with narrative briefing before manual backlog
▶Team Creation Wizard: AI pre-fills everything, humans override
▶Deploy Risk Assessment: visualized risk scoring before destructive actions

Specification Precision

Writing instructions so explicit that machines reproduce your exact intent, leaving no gaps for hallucination to fill.

▶21 agent specs with explicit capabilities, constraints, and output schemas
▶Per-agent tool scoping via toolsEnabled arrays (principle of least privilege)
▶Spec Pipeline: 4-level maturity system with LLM enrichment for agent-ready stories

Task Decomposition

Breaking complex objectives into agent-sized tasks. Knowing which subtasks an agent handles solo vs. which need human oversight.

▶5-stage intelligence pipeline: Discovery, Analyst, Prototype, Persona, Product Owner
▶Pointer-based handoff protocol with 40x token reduction vs. full-document passing
▶4 team topologies: Solo, Supervised, Collaborative, Hybrid
▶Agent execution pipeline: story, worktree, implement, test, PR, with clear handoff points

AI Systems

Context Architecture

Organizing knowledge so AI agents find exactly what they need. The "Dewey Decimal System" for your AI, covering persistent vs. ephemeral context.

▶Context Engine: getRelevantContext(), getRecommendationContext(), getStoryContext()
▶Tiered onboarding: 5 min (~70% AI effectiveness) to 30 min (~95%)
▶KB Quality Score: coverage assessment across 7 categories with gap identification
▶Scoped retrieval per feature, so consumer chat vs. internal agents see different context

RAG Pipeline Design

End-to-end retrieval-augmented generation: embedding strategy, hybrid search, prompt assembly, and output grounding.

▶Full embedding pipeline with pgvector, stale marking, and re-embedding triggers
▶Hybrid search: full-text + vector similarity with configurable weighting
▶Production-deployed Consumer Chat: external-facing RAG chatbot with scoped retrieval
▶Knowledge Chat: multi-turn RAG conversation with source attribution

Multi-Agent Orchestration

Designing how multiple AI agents collaborate: handoff protocols, shared state, error propagation, and recovery strategies.

▶AgentWeave: purpose-built orchestration engine with pipeline, solo, and collaborative modes
▶21 agents across 7 teams with defined topologies and supervisor styles
▶Training Arena: A/B testing different orchestration configurations against benchmarks
▶Team config export to Claude Code, Cursor, GitHub Copilot, Aider, and Continue.dev

Prompt Engineering at Scale

Moving beyond one-off prompts to systematic prompt management: database-stored, admin-editable, and multi-provider compatible.

▶21 agent prompts stored in database, editable from admin UI, not hardcoded
▶Model config table: prompts paired with per-feature model selection
▶Multi-provider compatibility: same patterns work across Anthropic Claude and Google Gemini
▶Structured output schemas: agents return typed JSON or Markdown with specific formats

Operations

Evaluation & Quality

Building systematic ways to measure whether AI output is actually correct, not just fluent. Catching the subtle failures.

▶Model Calibration: golden test cases with side-by-side comparison across providers
▶Training Arena: LLM-as-judge evaluation scoring completeness, precision, feasibility
▶Inference Inspector: full retrieval attribution to verify source relevance
▶Agent defect tracking with auto-creation on build/test failure

Trust & Security

Designing boundaries between human and machine. Building blast radius assessments. Ensuring AI systems fail safely.

▶8-scanner security pipeline: npm audit, auth guards, SQL injection, XSS, RLS, and more
▶Git worktree isolation for agent execution: disposable branches, never main
▶Per-agent tool scoping, team guardrails, and configurable review policies
▶RLS write-blocking on all 85+ tables for demo user, belt-and-suspenders protection

Cost & Token Economics

Calculating ROI of every AI operation. Right model for the right task. The new infrastructure cost optimization.

▶3-tier model config: Lite (Haiku/Flash), Standard (Sonnet), Advanced (Opus) per feature
▶Budget service: monthly spend tracking, burn rate, projection, stories-remaining
▶Per-agent token tracking logged to every run with cost estimation
▶AI Triage surfaces estimated token cost per recommended action

AI Observability

Monitoring AI systems in production, not just uptime, but output quality, cost trends, and retrieval attribution.

▶Inference Inspector: per-trace observability with plain-English explanations
▶Workforce Monitor: real-time activity, queue depth, token burn rate
▶Agent run logging: full prompt + response + metadata captured per execution
▶Security scanner with daily automated runs, auto-assignment, and auto-resolution

Key Design Decisions

Product decisions are more valuable than features. Each of these was a deliberate choice with a specific rationale and measurable outcome.

Pointer-based context passing

Why

Full-document passing bloats context windows and wastes tokens. Agents pass only UUIDs and summaries, forcing clean boundaries.

Outcome

40x token reduction in multi-agent pipelines while maintaining output quality.

Tiered onboarding tied to AI effectiveness

Why

AI effectiveness isn't a model problem. It's a context problem. Measuring context quality makes AI investment decisions tangible.

Outcome

Quantifiable framework: 5 min onboarding = ~70% AI effectiveness, 30 min = ~95%.

Recommendations before controls

Why

Traditional tools show a wall of data and expect humans to find the signal. AI-first means the system does the analysis and presents actions.

Outcome

AI Triage as default Work page, delivering a narrative briefing with cost estimates, risk, and recommended actions.

Anti-platform architecture (fork, don't rent)

Why

Multi-tenant SaaS locks customers into shared infrastructure with no customization. Each company should own their instance.

Outcome

Module manifest + schema split + config-driven agents, customizable without code changes.

See these skills in action

ProductIntel is live with a guided demo tour. No signup required.

Take the Demo TourView Capabilities