@upstate-web/uwc-skills
v0.3.0
Published
Curated Claude Agent Skills — 22 production-validated skills covering AI cost guards, prompt engineering, extraction, scoring, testing, security, billing, infrastructure, SaaS patterns, and more.
Maintainers
Readme
@upstate-web/uwc-skills
Curated Claude Agent Skills extracted from Upstate Web Co.'s production pipeline — cost guards, prompt engineering, AI extraction, weighted scoring, memory systems, and three-gate shipping for AI-native apps.
What this is
A bundle of battle-tested Claude Agent Skills pulled from real production projects. Every skill in this package has shipped to production at least once. Each encodes a specific engineering pattern — what problem it solves, the exact code to drop in, and the regressions it prevents.
v0.2.0 ships six skills, all extracted from production AI pipelines across multiple client projects:
| Skill | What it does | Stacks |
|---|---|---|
| ai-cost-guard | Per-tenant daily cost cap + immutable call log for any LLM API | Django, Astro / Cloudflare D1 |
| weighted-scoring-engine | Transparent weighted scoring with Decimal-drift guard + factory pattern | Django, generalizable |
| three-gate-pipeline | Prototype → MVP → Launch client sign-off gates for app projects | Any agency / services workflow |
| claude-prompting | NEW — The meta-skill for prompt engineering. System=contract, User=data. Four prompt components, caching rules, versioning discipline, v2-promotion protocol | Any LLM project |
| ai-extraction | NEW — Messy text → structured rows pipeline. Five SYSTEM_PROMPT rules, merge-don't-overwrite logic, dedup patterns, confidence scoring | Django, Python, generalizable |
| memory-system | NEW — Cross-session agent memory architecture. MEMORY.md index, topic files, archive lifecycle, truncation danger prevention | Any AI coding agent project |
Install
npm install @upstate-web/uwc-skillsThe package contains no runtime JavaScript — just markdown documents. Installing places the skills/ directory under node_modules/@upstate-web/uwc-skills/skills/, where Claude Code and MCP-aware tools can load them.
Use with Claude Code
Each skill is a directory with a SKILL.md (Anthropic Agent Skills format: YAML frontmatter + body). Point Claude Code at whichever you want active:
# From a project root, link a single skill into .claude/skills/
ln -s node_modules/@upstate-web/uwc-skills/skills/ai-cost-guard .claude/skills/ai-cost-guardOr copy all six:
cp -r node_modules/@upstate-web/uwc-skills/skills/* .claude/skills/Claude will load the SKILL.md on session start and trigger the skill when its description matches the task at hand.
Use with MCP
The companion @upstate-web/uwc-pipeline-mcp server exposes these skills as MCP tools (list_skills, get_skill). That package depends on this one — installing the MCP server is the quickest way to make these skills available across every MCP-compatible client.
Why these skills
UWC runs a full intake → scope → build → governance → launch → retro pipeline. The six skills in v0.2.0 are the ones that generalize best beyond UWC's specific stack:
ai-cost-guardis the lowest-blast-radius skill to add to an existing codebase and the highest-regret one to skip. A single runaway loop has bankrupted more than one prototype.claude-promptingis the meta-skill every prompt engineer needs but nobody writes down. The "System = contract; User = data" rule alone prevents the three most common prompt engineering failures.ai-extractionturns the black art of LLM extraction into a repeatable pipeline with five mandatory rules and an edge case catalog.weighted-scoring-engineencodes the quiet Decimal-drift bug. Three equal 10-weights × three perfect scores of 100 will produce99.99…8unless you.quantize()before save.memory-systemprevents the silent truncation that loses agent context across sessions.three-gate-pipelineturns "vibe-driven client engagements" into "billable sign-off milestones."
Versioning
- Patch (
0.2.x) — typo fixes, clarifications, bumped internal skill versions with no trigger change. - Minor (
0.x.0) — new skills added, new skill versions with behavior changes, trigger broadening. - Major (
x.0.0) — skill renames, breaking trigger changes, or directory-structure changes.
Each individual skill also carries a version field in its frontmatter (v1.0, v1.1, etc.) that bumps independently.
Contributing
This is an opinionated bundle — every skill extracted from a production project, every version bump traced to a retro. External PRs welcome for typos, clarifications, and portability fixes.
New skills enter the bundle only if they pass three tests:
- Portability. The skill works outside UWC's specific stack — no hardcoded references to private paths or specific tooling.
- Leakage. No live links to private repos, no admin schema details, no API keys or shared secrets.
- Usefulness. A developer using Claude Code on their own project gets meaningful value from the skill without needing any other tooling.
Skills that don't pass all three tests are dropped, not deferred.
Related
- UWC blog: upstate-web.com/blog
- Agent Skills spec: docs.anthropic.com/en/docs/agents/skills
- MCP spec: modelcontextprotocol.io
- Companion MCP server: @upstate-web/uwc-pipeline-mcp
- D1 read-only MCP: @upstate-web/d1-readonly-mcp
License
MIT. Use these in commercial work freely; credit appreciated, not required.
