{"ok":true,"count":14,"skills":[{"id":"code-python","track":"code","title":"Python for agent-built services","tokens":"~190","why":"Frameworks make whole mistake classes unwritable, so the agent spends tokens on the task instead of on validation and plumbing.","prompt":"Create and install a skill named `code-python` (Python for agent-built services). It instructs the agent as follows: Write Python services with FastAPI + Pydantic: one model drives routing, validation, and the OpenAPI schema — never hand-write type checks. One SQLModel class serves as both table and validator. CLIs are one Typer-decorated function. Type everything; let the framework and tests catch errors, not code review. Between Python-only services reuse the same Pydantic models; when sibling services aren't Python, use gRPC between services and keep FastAPI at the edge.","checks":["all I/O typed with Pydantic models","no hand-written validators","tests cover happy path + one edge"]},{"id":"ui-mobile","track":"ui","title":"Mobile: FastAPI + HTMX + Tailwind + Capacitor","tokens":"~200","why":"One language and zero client state means the agent writes every feature once and it ships to both app stores.","prompt":"Create and install a skill named `ui-mobile` (Mobile: FastAPI + HTMX + Tailwind + Capacitor). It instructs the agent as follows: Build mobile apps as server-rendered HTML: FastAPI routes return fragments, HTMX attributes (hx-get/hx-target) give SPA behavior with near-zero JavaScript, Tailwind utilities handle responsive layout inline, Capacitor wraps the mobile viewport into iOS/Android shells. No client state, no fetch calls, no Swift/Kotlin.","checks":["zero hand-written fetch calls","no client-side state store","layout verified at 360px width"]},{"id":"code-token","track":"code","title":"Token-efficient code","tokens":"~190","why":"Every token spent writing or re-reading code is a token not spent on the task; cheap code compounds across sessions.","prompt":"Create and install a skill named `code-token` (Token-efficient code). It instructs the agent as follows: Write code cheap to generate and cheap to re-read: type everything, one canonical pattern per job (FastAPI route, Pydantic model, Typer command). Early returns, table-driven branches, JSON-first data. Delete dead code — every dead line taxes every future read.","checks":["no nested depth >2","no dead exports","single pattern per job"]},{"id":"code-secure","track":"code","title":"Secure by default","tokens":"~160","why":"Insecure code fails late and expensively; validated boundaries fail fast on the first bad input.","prompt":"Create and install a skill named `code-secure` (Secure by default). It instructs the agent as follows: Validate every input at the boundary (type, length, allowlist). Parameterize queries, escape HTML output, never eval(). Secrets only in env, never in code or logs. Reject oversized payloads (>100KB) before parsing.","checks":["all inputs validated","no eval/innerHTML with raw input","no secret in repo"]},{"id":"code-boundaries","track":"code","title":"Service boundaries","tokens":"~220","why":"The wrong protocol between services costs speed forever; the right one is a one-line decision made once.","prompt":"Create and install a skill named `code-boundaries` (Service boundaries). It instructs the agent as follows: Connect services by rule, not habit: same Python system → shared Pydantic models over HTTP; other languages or hot internal paths → protobuf plus gRPC, FastAPI stays at the edge; events and queues → Pydantic-validated messages. Build the whole app in Python by default, but carve out other languages where Python technically fails: single-digit-millisecond budgets use Go or Rust, over 100k live connections use Erlang or Elixir, heavy offline clients and games and widgets go native.","checks":["protocol chosen by rule, not default","non-Python parts named with their technical reason","FastAPI at the edge in mixed systems"]},{"id":"ui-shell","track":"ui","title":"UI shell in 40 lines","tokens":"~240","why":"Humans get the interface they prefer; metadata makes it legible to agents without dumbing it down.","prompt":"Create and install a skill named `ui-shell` (UI shell in 40 lines). It instructs the agent as follows: Build mobile-first: design at 360px width, then scale up — full-monitor layouts only when the task genuinely needs a big screen. Semantic HTML shell (
/
/