Design mentor

I wanted more than "an assistant that answers" — I wanted a mentor at the Head of Product Design level: one who asks leading questions, praises only when deserved, names the framework, and offers two or three options instead of a single "correct" truth. One I could summon with "call the mentor" in any project.

I built that agent and published it as an open repository: github.com/mmkeewa/design-mentor.

Architecture: 4 layers + an orchestrator

The 26 skills aren't a flat list — they're layered the way a real Head thinks:

• Orchestrator — discovery-delivery, a router over the whole cycle. It decides "what is needed now" — strategy, research, ideation, or polish — and calls the right skill.
• Reference — visual literacy (visual-literacy) and platform sources of truth (platform-guidelines): Apple HIG, Material 3, WCAG, canonical design systems.
• Flow — 15 skills mapped to product-cycle phases: from product-strategy and ux-research through clean-ui, design-systems, accessibility to handoff-dev-collab and design-impact.
• Domain — cross-cutting lenses fintech-ux, ecom-ux, edtech-ux. They overlay any phase: money, carts, and learning each have a different cost of error.
• Ops — the leadership side: team management, hiring, stakeholders, presentations, design ops.

Every skill carries boundaries with its neighbours ("when to call someone else, not me"), common traps, sources with URLs, and the mentor's answer template: praise → diagnosis → framework → 2–3 options → next step.

Two hard-wired principles

Trend vs fashion. The mentor speaks the industry's language but won't drag every fad into fintech — any fashionable pattern passes a four-question gate first: does it solve the task, does it hurt accessibility, does it fit the domain, will it still make sense in six months.

Living skills verify against primary sources. Seven skills are marked living — their sources (guidelines, regulation, galleries) go stale fast. Once a week, GitHub Actions runs a script that reads the URLs from each living skill's frontmatter, compares content hashes against the previous run, and commits a STATUS.md saying which skills need re-checking.

Two more features

The mentor knows my GitHub. At the start of a session it checks my profile via the API, and when a separate skill repository fits the user's task (say, design-discovery for reference gathering), it recommends it with a one-line install command. When I add a new repo, the mentor finds it on its own — no agent edits needed.

One-command install. install.sh lays out the agent and skills as symlinks into ~/.claude — a git pull immediately propagates fresh content. It ships with --check, --dry-run, --uninstall, and --update.

The result

• The "call the mentor" trigger works in any Claude Code session
• 26 skills, each a self-contained document with sources
• Weekly automated freshness checks for living skills
• Live recommendation of my external skills via the GitHub API