MMCPMarkets
MV

myatminlu/vector-skills

Developer tools
20 stars Quality 65 Trend 18MIT
Repository Install

Claude Code skills library — senior NestJS/Node.js backend engineer, and more

Overview

A personal library of Claude Code skills. Each skill teaches the agent a domain — how to think, what rules to follow, what mistakes to avoid — when working on that kind of code. (More skills planned — react-frontend-guidelines, expo-mobile-guidelines, etc.) Or manually — skills live in ~/.claude/skills/ so Claude Code auto-discovers them: Next Claude Code session will load the skill automatically. No restart or install step. Edit the files in place. Changes take effect on the next Claude Code session. A skill is a folder with at least a SKILL.md file containing YAML frontmatter (name, description). See any existing skill as a template. - Keep SKILL.md scannable (≤ 500 lines). Deep dives go in references/. - Each reference file: TL;DR → rules (good/bad code) → anti-patterns → review checklist. - Write description in the frontmatter so Claude triggers the skill reliably (concrete keywords + "use when X" framing). Max 1024 characters. - Add evals/evals.

README

vector-skills

A personal library of Claude Code skills. Each skill teaches the agent a domain — how to think, what rules to follow, what mistakes to avoid — when working on that kind of code.

Skills in this collection

Skill What it does
nestjs-dev-guidelines Senior NestJS / Node.js backend engineer: execution discipline, folder structure, naming, module design, code quality, decision trees, API design, standard responses, pagination/filters/sorting, validation, error handling, exception filters, security, auth/RBAC, multi-tenancy, DB design + ORM patterns, migrations, cascade rules, pipes/interceptors/guards, events, background jobs, configuration, logging, observability/tracing, testing, performance, N+1 elimination, caching, Swagger docs, AI product patterns (LLM gateway, SSE streaming, usage metering & cost), code review checklist + anti-patterns, modern Nest stack, health/readiness/shutdown, source-of-truth freshness checks, webhooks, file uploads, decorators/scopes/dynamic modules, DDD layered/hexagonal architecture (ports/adapters/use cases). SKILL.md + 43 granular reference files + evals.
code-implementation Architect-grade workflow for changing code in an existing codebase — features, fixes, refactors, integrations, migrations. Read the real code before writing any, trace the root cause instead of patching symptoms, reuse what exists instead of adding near-duplicates, verify end to end. SKILL.md + 5 reference files (investigation, root-cause, design-review, code-quality, verification) + evals.
bug-fixing Disciplined defect workflow — reproduce, read the real code, find the true root cause, design the minimal correct fix at the right layer, verify it, then hunt for clones of the same defect. Covers crashes, wrong output, flaky tests, regressions, races, leaks, perf degradation, works-locally-not-in-prod. SKILL.md + 3 reference files (root-cause techniques, anti-patterns, report templates) + evals.
code-quality-check Read-only, evidence-backed quality review producing severity-ranked findings, a scorecard, and a merge/ship verdict. For PR/diff review, pre-merge gates, and second opinions on agent- or teammate-written code. Hands off to the implementation workflow rather than writing the fix. SKILL.md + 6 reference files + scripts/repo_signals.py + evals.
project-structure-refactoring Cross-language project structure and structural refactoring: screaming-architecture / feature-first layout principles, safe staged restructuring (inventory → mapping table → approval gate → green stages), git mv history preservation, god-file splitting (symbol-level mapping, shrinking facades, extraction order), silent-breaker census (CI path filters, CODEOWNERS, Dockerfiles, dotted-path strings), and per-stack reference layouts for JS/TS (Node, React, Next.js, Expo), Python (src layout, FastAPI, Django), Go (cmd/internal), and monorepos (apps/packages, workspaces). Structural only — hands code-level refactoring to code-implementation. SKILL.md + 7 reference files + evals.
codebase-audit Forensic whole-codebase audit and staged refactor plan — maps the system, verifies features are actually wired end to end, finds duplicated services and copy-pasted modules, band-aid fixes, dead code, and legacy-vs-new implementations living side by side. Emits a numbered report set. SKILL.md + 4 reference files + evals.

(More skills planned — react-frontend-guidelines, expo-mobile-guidelines, etc.)

Install

Via the skills CLI (recommended):

npx skills add myatminlu/vector-skills

Or manually — skills live in ~/.claude/skills/ so Claude Code auto-discovers them:

git clone  /vector-skills
ln -s /vector-skills/skills/nestjs-dev-guidelines \
      ~/.claude/skills/nestjs-dev-guidelines

Next Claude Code session will load the skill automatically. No restart or install step.

Updating a skill

npx skills update

Editing a skill

Edit the files in place. Changes take effect on the next Claude Code session.

cd /vector-skills/skills/nestjs-dev-guidelines
vim references/08-pagination-filters-sorting.md

# optional structural validation
python3 ~/.claude/skills/skill-creator/scripts/quick_validate.py .

git add . && git commit -m "tighten pagination rule"
git push

Adding a new skill

cd /vector-skills/skills
mkdir my-new-skill
cd my-new-skill
# Create SKILL.md (required) + optional references/, evals/, assets/, scripts/
ln -s $PWD ~/.claude/skills/my-new-skill
git add . && git commit -m "Add my-new-skill"

A skill is a folder with at least a SKILL.md file containing YAML frontmatter (name, description). See any existing skill as a template.

Authoring conventions

  • Keep SKILL.md scannable (≤ 500 lines). Deep dives go in references/.
  • Each reference file: TL;DR → rules (good/bad code) → anti-patterns → review checklist.
  • Write description in the frontmatter so Claude triggers the skill reliably (concrete keywords + “use when X” framing). Max 1024 characters.
  • Add evals/evals.json with 5–10 substantive test tasks to catch regressions.

Running evals

Each skill ships evals/evals.json (prompt, expected output, must_include keyphrases) and a coarse grader. Run the prompts against an agent with the skill loaded, save the replies as a JSON list of { "id": N, "response": "..." }, then grade:

python3 skills//evals/scripts/grade.py responses.json

It is a substring check, not an LLM judge — it catches gross misses (wrong topic, missing required terms), not subtle correctness. Exit code is non-zero if any eval fails; pass --threshold 0.8 to allow partial misses.

Author

Built and maintained by Myat Min Lu — a developer curating a personal collection of Claude Code skills distilled from day-to-day engineering work. Each skill captures the conventions, rules, and hard-won lessons worth teaching an agent once so they don’t have to be re-explained every session.

Contributions, forks, and feedback are welcome.

License

MIT — feel free to fork and adapt.

View the full README on GitHub

Recommended Tools

Try a different keyword or remove a filter.

Install

npx skillfish add myatminlu/vector-skills
New to MCP? Start with the mcpskill.org guide →