skills/

Agent skills extracted from production work on autoraters for UX and product evaluation, and from darrenhead.com itself.

A skill is a single Markdown file (SKILL.md) with YAML frontmatter. Supported agents (Claude Code, Cursor, Codex, Cline, Windsurf, OpenCode, Aider, and 40+ others via the skills CLI) auto-load installed skills and invoke them when a session's content matches the skill's description trigger. Adopters get production-grade scaffolding (templates, decision frameworks, anti-patterns) in their own agent sessions without inventing the methodology themselves.

What's published

Skill	One-liner	Provenance
autorater-rubric	Battle-tested rubric structure for LLM-as-judge evaluations: SxS vs SSE selection, calibration loop, agreement thresholds, anti-patterns.	Production autorater work
persona-aware-disclosure	Adapt the same factual answer for recruiter / engineer / executive audiences without losing signal or branch-sprawling.	darrenhead.com chat surface
multimodal-structured-extraction	Image → zod-typed JSON via a vision LLM. Schema-first design, validation gates with retry, locale handling, anti-patterns.	Shipping multiple image-to-JSON pipelines
reddit-pain-mining	Validate a SaaS idea against real Reddit pain-point threads. The full niche → search → score → cluster → one-page brief pipeline.	Pico Pitch (open source)
genux-patterns	When an AI answer IS a UI, not text about a UI. Decision tree, layout primitives, validation gates, graceful fallback. (v0.1 — deepens with production source)	Public GenUX surfaces + shipped chat work

The companion catalog page on the site lives at darrenhead.com/skills and surfaces install commands + source links per skill.

Install

If you have the skills CLI:

npx skills add darrenhead/darrenhead.com/skills/<slug>

For example:

npx skills add darrenhead/darrenhead.com/skills/autorater-rubric

The CLI auto-detects whichever agent you've configured and drops SKILL.md into the right location (~/.claude/skills/<slug>/SKILL.md for Claude Code, .cursor/rules/ for Cursor, etc.) or use -a <agent> to target one explicitly. The skill becomes available in your next session — either via implicit trigger (the agent auto-pulls when your conversation matches the description keywords) or explicit invocation (/<slug>).

If you don't have the CLI yet, the manual fallback:

mkdir -p ~/.claude/skills/autorater-rubric
curl -L https://raw.githubusercontent.com/darrenhead/skills/main/skills/autorater-rubric/SKILL.md \
  -o ~/.claude/skills/autorater-rubric/SKILL.md

Philosophy

These skills publish methodology, not implementation. The thinking — when to use SxS vs SSE, why centered odd-numbered scales, how to write dimension anchors, the calibration loop, the anti-patterns experience teaches — is what generalises. The specific production prompts and tunings stay inside the systems they were built for; what's published is the framework an adopter needs to build their own version, tailored to their model and their dimensions.

The point is recursive: skills extracted from work that ships → installed by other AI builders → their work improves → the loop closes. If a skill helps you ship something, @darrenhead wants to hear about it.

Contributing back

If you adapt one of these skills and ship something with it, an open issue / PR on this repo with the "in the wild" entry is welcome. The skills are MIT-licensed; the only ask is provenance — keep the link to this repo so others can find the source.

What's coming

Roadmap (no dates — these ship when they're ready):

• grounded-corpus-citations — pgvector + Gemini embeddings + the citation-back pattern darrenhead.com uses for "every claim links to source"
• anti-abuse-quotas — per-cookie quotas, rate limits, BotID — the defensive boilerplate every public AI app needs
• chat-first-streaming-ui — AI SDK streaming + seed-prompt + persona-aware composer patterns
• menu-multimodal-translate — the Taberu pipeline (image → 29-language menu translation preserving culinary context)
• ja-receipt-ocr — the Denpyo pipeline (Japanese receipt → vendor/date/amount/消費税/category)

Vote with installs — the ones that get pulled most get fleshed out first.